← PART ONE: WHO IS ROCKEFELLER?

THE LONG READ · PART TWO · OCTOBER 31, 2026

If AI is the new oil, part two: rebates, rockets, and robots

Space data centers, robotaxis, Meta’s quiet integration, Google’s counterattack, the volume discounts that made Rockefeller, and the boom-bust cycle in compute. Seven hypotheses, each checked against the evidence.

~60 MIN READREADS ON ITS OWNFIGURES AS OF OCT 31, 2026NOT INVESTMENT ADVICE
LIVE: THE HYPOTHESIS TRACKER — ARE THE BETS MOVING? →

Part One argued that in oil, value flowed to whoever held the scarcest step, and that step kept moving. This part asks what happens when you push on the castings.

The first essay settled on a working map. Jensen Huang looked most like Vanderbilt, collecting a toll on everything that moves. Musk looked like a rail baron crossed with the Nobel brothers, not like Rockefeller. Google was the most Standard-like integrator. Open-weight models were Spindletop. Texas, auditing data centers before they could plug in, was acting like the Railroad Commission again. And the largest prize, as with the car, would probably go to whoever used cheap intelligence to do something new.

A map like that invites objections. Some are about players the first essay underweighted: Meta, the robot makers, a SpaceX that now rents compute to its rivals. Some are about the plot: who really gets the rebates, who is overplaying their hand, and whether the whole industry is headed for an oil-style cycle of glut and shortage. This essay collects seven of those objections, states each as a hypothesis, and tests it against the evidence available at the end of October 2026.

Seven hypotheses

  1. Orbit. Musk’s space data centers are a serious potential outcome, launch becomes the new railroad, and the advantage would last much longer than most AI positions, more like Nvidia’s than like the OpenAI-versus-Anthropic race.
  2. Robots are the new cars. Robotaxis and robots are AI’s gasoline moment, underweighted by everyone but a small group of technologists, with Tesla and SpaceX best positioned and Google’s Waymo holding the technology but less commercialized.
  3. Meta was underweighted. Zuckerberg has navigated every platform shift in two decades, owns his infrastructure and models, and runs the clearest AI-for-ads flywheel. And maybe the future interface is closer to Neuralink than to glasses.
  4. Google is Standard Oil after its prime. OpenAI is best positioned to dominate AI advertising and could erode Google substantially.
  5. The map is upside down. Hyperscalers and neoclouds are the railroads that crave guaranteed volume. The labs that control volume are Standard and get the rebates. Someone will consolidate the neoclouds the way Standard rolled up refiners, and Oracle is the Pennsylvania Railroad overplaying its hand.
  6. Revenue breadth is survival. Companies with many revenue streams outlast single-source players; pure token production loses value fast; debt-financed neoclouds are the most fragile.
  7. Booms and busts. AI follows an oil-style cycle: breakthroughs cut prices, cheap prices unlock demand, demand pushes prices back up. Consumer spending settles around $50 to $100 a person a month. Amazon could win agentic shopping if it gets its act together. And Nvidia, secure in chip design, can move across the stack.

Each gets a verdict: where the evidence supports it, where it doesn’t, and where it’s too early to say. The figures are dated and sourced; the research trail is at the bottom. Where something is judgment rather than fact, it is labeled as analysis.

The short version: volume buys rebates, breadth buys time, and orbit is a real option, not yet a railroad.

A note on structure. Part One used “Part One/Two/Three” as section banners. To avoid confusion, this essay’s sections are numbered chapters. Chapters 1 through 5 test the hypotheses. Chapter 6 steps back to survey the players almost nobody is watching, and a closing synthesis scores everything.

CHAPTER 1The railroad to orbit

Part One treated SpaceX’s orbital data-center plan as a footnote, flagged by the company’s own filing as “novel and untested.” The hypothesis here is that this undersold it. Musk’s space data centers deserve to be treated as a serious potential outcome. If they work, launch becomes the new railroad. And the advantage could last much longer than most positions in AI: more like Nvidia’s hold on chips than like the months-long leapfrogging between OpenAI and Anthropic.

That deserves a careful hearing, starting with Musk’s own case, in his own words.

Musk’s case: the power wall

The fullest statement of the argument came in a long interview with Dwarkesh Patel and Stripe’s John Collison, published February 5, 2026. Musk’s starting point is not rockets. It’s electricity.

“The output of chips is growing pretty much exponentially, but the output of electricity is flat.”— ELON MUSK, DWARKESH PODCAST, FEB 5, 2026

On Earth, he argued, the binding constraint on AI is power plants, transformers and permits, not chips. “The turbines are sold out through 2030,” he said, and SpaceX and Tesla might have to make their own turbine blades. At Davos in January he had put it more simply: “I think the limiting factor for AI deployment is fundamentally electrical power,” adding that before long, “we’ll be producing more chips than we can turn on.”

Space, in his telling, removes the constraint. There are no nights, seasons or clouds in the right orbit, and no atmosphere absorbing sunlight. “Any given solar panel can do about five times more power in space than on the ground,” he said, and because the sun never sets, “it’s 10 times cheaper because you don’t need any batteries.” Then came the timeline that made headlines:

“My prediction is that it will be by far the cheapest place to put AI. It will be space in 36 months or less. Maybe 30 months.”— ELON MUSK, DWARKESH PODCAST, FEB 5, 2026

The scale he described was enormous: “at least, five years from now, a few hundred gigawatts per year of AI in space and rising,” and perhaps “around a terawatt a year” before rocket-fuel supply becomes the limit. For comparison, Dwarkesh noted that the whole world has maybe 20 to 25 gigawatts of AI compute today. Asked how many Starships it would take to fly 10,000 launches a year, roughly one an hour, Musk said “as few as 20 or 30.” On chips, he floated a “TeraFab” and admitted, “I don’t know how to build a fab yet. I’ll figure it out.” His biggest supply worry, he said, was memory.

What SpaceX has actually built

Strip away the projections and the company behind them is now public and reporting. Its third-quarter 2026 results, released October 20, make the most important point of this chapter before orbit even enters the picture.

SPACEX BY SEGMENT, Q3 2026SpaceX Q3 2026 earnings release (SEC); customer names via NBC News
SegmentQ3 2026 revenueQ3 2025Notes
Space (launch, Starship)$1.77B$1.67BOperating loss of $7M; first revenue Starship flight
Connectivity (Starlink)$5.25B$3.10B13.6M consumer subscribers; ARPU $69/month, down from $83
Super Intelligence (xAI, compute)$7.56B$0.81BMostly renting compute to Anthropic, Google and Reflection AI; first operating profit ($431M)
Total$14.6B$5.6BNet income $4.4B; Q3 capex $19.6B, of which $16.4B for AI

The biggest line in SpaceX’s income statement is no longer rockets or Starlink. It’s the segment it calls Super Intelligence, which grew more than ninefold in a year to $7.56 billion. According to NBC’s reporting, that revenue came mostly from renting data-center capacity to Anthropic, Google and Reflection AI. NBC put it pointedly: the business looks “closer to a cloud provider’s than an AI lab’s.” SpaceX had 1.7 gigawatts of nameplate compute at the end of September, up from 0.6 a year earlier. It said it would reach 2.5 to 3 gigawatts by the end of the year and expects fourth-quarter capex of $35 to $40 billion. It also completed the acquisition of Cursor, the coding tool, at an implied $60 billion.

On launch, SpaceX’s position is remarkable. Its prospectus says it has launched more than 80 percent of the world’s mass to orbit every year since 2023: 2,213 metric tons and 170 launches in 2025 alone. Musk claimed almost 90 percent on the third-quarter call, with China at about 6 percent and everyone else about 4. Independent counts differ: BryceTech counted 165 SpaceX orbital launches in 2025, 51 percent of the global total by number. Either way, no one is close. Starship reached orbit for the first time on September 28, 2026 and deployed 26 next-generation Starlink satellites, though one engine failed on ascent and the flight was cut short. On-orbit refueling, which deep-space plans and much of the cost story depend on, has “not yet” been “demonstrated or attempted,” the prospectus says. SpaceX plans to try it early in 2027.

The orbital compute plan itself has a name and a supplier. In August SpaceX said its Starmind satellites will carry Nvidia Rubin GPUs and Vera CPUs. Musk called it “essentially an optimized Vera Rubin NVL72 computer” and said “we’re exclusive to NVIDIA.” The dates already conflict. In August, Musk said launches would start “next year.” In October, President Gwynne Shotwell said Starmind would begin flying in 2027. The prospectus says deployment could begin “as early as 2028.” The FCC is still reviewing SpaceX’s January application to launch up to one million satellites for what it calls the “SpaceX Orbital Data Center system.”

The case against: heat, servicing, and the 15 percent problem

The best counterarguments come from the same interview and from a companion essay Dwarkesh published that day, “Notes on Space GPUs,” which he opens with a disarming warning: “Assume all the numbers are wrong.” The objections fall into five groups.

1. Energy is a small share of the bill. This is the deepest objection. Dwarkesh put it to Musk directly: “only 10-15% of the total cost of ownership of a data center is energy. That’s the part you’re presumably saving by moving this into space. Most of it’s the GPUs.” In his notes he puts chips at roughly 70 percent of total cost. SpaceX’s own prospectus concedes the point: “We do not believe that moving compute to space in and of itself will have a meaningful impact on the cost of procuring processors.” Orbit can at best attack the smaller part of the cost stack, and only if launch is cheap enough not to add a new big one.

2. Heat. “In space, you can’t convect heat away, because there’s no air.” Every watt of compute needs roughly a watt of solar collection and a watt of radiator capacity. Dwarkesh’s rough estimates put a whole orbital system at about 85 watts per kilogram, which turns into a lot of mass to launch. Musk’s answer was physics: run hotter, and “if you increase the operating temperature by 20% in degrees Kelvin, you can cut your radiator mass in half.” The constraint is real today. Google’s first Suncatcher prototype, launched October 1, 2026, runs its model in 15-minute stretches because of heat limits.

3. Servicing and depreciation. GPUs fail. Dwarkesh cites Meta’s Llama 3 paper, in which a 16,000-GPU cluster saw a failure roughly every three hours. On the ground, a technician swaps the part. In orbit, SpaceX says plainly that “we do not anticipate servicing or repairing processors in space.” Musk argued that GPUs are now reliable once early failures are screened out on the ground. One independent analysis assumes about 9 percent of chips fail each year. Either way, a satellite full of chips that will be obsolete in a few years is a depreciation bet that can’t be upgraded.

4. Launch cost. Estimates of today’s cost range from about $1,500 to $2,500 a kilogram depending on assumptions. Dwarkesh cites around $2,500 for Falcon 9; a Forethought analysis uses about $1,500. Forethought finds that orbital data centers reach full cost parity with grid-connected ones at around $100 a kilogram, roughly ten Starship reuses. Google’s Suncatcher paper projects launch below $200 a kilogram by the mid-2030s. The engineer Andrew McCalip’s widely cited calculator put a one-gigawatt orbital system at about three times the cost of a terrestrial one over five years. His own first estimate had been seven to ten times.

5. Latency, bandwidth and training. Orbit adds up to about 50 milliseconds of latency, which matters little for an agent grinding through a long task. Bandwidth is the harder problem. Starlink’s lasers run at about 100 gigabits per second, against 400 for a terrestrial InfiniBand link. Dwarkesh concludes that pretraining likely stays on the ground and space, if it works, suits inference.

The people building rivals are cautious too. Google’s Suncatcher lead, Travis Beals, said: “I don’t see this being something where it’s cheaper to do this in the next five years.” Jeff Bezos has said orbital data centers could beat Earth-based costs in 10 to 20 years. And SpaceX’s own prospectus contains the most interesting risk factor in the whole filing: if “advances in nuclear energy, significantly reduce energy costs or alleviate infrastructure constraints, the viability of our orbital AI compute infrastructure may be materially diminished.”

The counter-counterpoint is that Musk is right about the ground. Wood Mackenzie data cited by Power Engineering show about 110 gigawatts of gas-turbine orders at the end of 2025 against 60 to 70 gigawatts a year of global manufacturing capacity. Lead times run about six years, and prices are heading toward $600 per kilowatt by the end of 2027, almost triple 2019. Part One documented Texas auditing data centers before letting them connect. The power wall is not imaginary. The question is whether it is cheaper to climb or to go to orbit.

ANALYSISThe honest read is that orbit is a real option with a long fuse, and the timelines disagree with each other even inside SpaceX. Musk said 30 to 36 months to cheapest in February; Musk said Starmind launches in 2027 in August; the prospectus says deployment as early as 2028; Google’s project lead says not cheaper within five years; Bezos says 10 to 20. The physics doesn’t forbid it. The economics depend on launch costs falling roughly another order of magnitude and on chips surviving unserviced long enough to pay back.

Who goes up, and how it plays out

The field is no longer one company. Google’s Project Suncatcher, announced in November 2025, flew its first prototype with four TPUs and the satellite company Planet in October 2026. It launched on a SpaceX rideshare. Starcloud, backed by Nvidia, flew a single H100 in November 2025, raised $250 million at a $2.3 billion valuation in August 2026, and slipped its second satellite to 2027. Blue Origin filed with the FCC in March 2026 for more than 50,000 orbital data-center satellites, “Project Sunrise.” Its New Glenn rocket landed a reused booster for the first time in April, then stranded a customer satellite because of an upper-stage problem. Aetherflux plans a first data-center satellite in early 2027. China is furthest along outside SpaceX: ADA Space and Zhejiang Lab launched 12 computing satellites in May 2025 and a second batch on October 14, 2026, testing in-orbit training, liquid cooling and deployable radiators. In January 2026, ADA said it had run Alibaba’s Qwen3 model in orbit.

Then there are the governments. The FCC received nearly 1,500 comments on SpaceX’s million-satellite filing. Amazon petitioned to deny it as a “lofty ambition rather than a real plan.” The American Astronomical Society called it “an unprecedented threat to the visual integrity of the night sky.” The FCC chair, Brendan Carr, publicly brushed off Amazon’s petition. Low-Earth-orbit satellites must deorbit within five years of mission end under the FCC’s 2022 rule. In July 2026 the FCC found it “premature” to adopt constellation-level debris standards. The Office of Space Commerce has opened a voluntary pilot certification for novel activities including orbital data centers. Under the Outer Space Treaty, states remain responsible for their companies’ activities in orbit. And data-sovereignty scholars point out that compute in orbit can sit outside the reach of national data-localization laws, which is exactly why governments will not leave it alone.

SPACE DATA CENTERS: SIX WAYS IT COULD GOAnalysis; facts as of Oct 31, 2026 from filings and reporting
ScenarioWhat has to be trueWho’s involvedWho winsOil-era parallelSignals to watch
1. Musk on scheduleStarship reaches full reuse and weekly cadence in 2027; Starmind compute satellites (Nvidia Vera Rubin) fly from 2027–28; launch falls toward ~$100/kgSpaceX/xAI, Nvidia (exclusive chip supplier), Tesla (solar cells), anchor tenants like Anthropic and GoogleSpaceX becomes railroad and refinery at once, with a cost floor nobody else can match for yearsStandard’s pipelines: lower cost per barrel than any rival’s routeRefueling demo (planned early 2027); Starship cadence; first Starmind launch
2. The launch toll roadOrbital compute works, but mass to orbit is the bottleneck; rivals need launch they don’t ownGoogle Suncatcher (its prototype flew on a SpaceX rideshare), Starcloud, Aetherflux, Blue Origin’s Project Sunrise, Chinese constellationsWhoever owns cheap, frequent launch. Today that is SpaceX, at 80%+ of the world’s mass to orbitThe trunk railroads before pipelines: everyone pays the tollRivals’ launch contracts; New Glenn reliability; Chinese reusable rockets
3. Railroad Commission in orbitGovernments ration orbital shells, spectrum and debris risk; national-security rules shape who may launch whatFCC (1M-satellite application still under review), Office of Space Commerce, ITU, astronomers, Amazon, China, Outer Space Treaty statesIncumbents with licenses and spectrum; politically connected operatorsTexas Railroad Commission proration; the Interstate Commerce Act forcing railroads to serve all shippersFCC ruling on SpaceX’s orbital data-center filing; constellation-level debris rules; any common-carrier demands for launch
4. Niche, not nexusSpace works for some inference, defense and sovereign workloads but never becomes the main place AI runs; training stays on the groundGovernments, defense agencies, SpaceX, Starcloud, China’s Three-Body constellationSpecialists; SpaceX still sells launch and connectivityKerosene for lighthouses: real, profitable, smallShare of new compute capacity in orbit (Forethought expects under ~10% before 2030)
5. Never cheaper than the groundTurbines, nuclear, solar-plus-storage and grid build-outs catch up; launch costs stall; servicing and heat stay hardUtilities, GE Vernova and turbine makers, nuclear developers, terrestrial neocloudsGround landlords. SpaceX keeps a large terrestrial cloud business it already hasPipelines that were never built because rail rates fellTurbine lead times; nuclear deals; SpaceX’s own risk factor that cheaper terrestrial energy could undercut orbit
6. Two orbitsThe U.S. and China each build sovereign orbital compute; export controls extend to space hardwareSpaceX, ADA Space/Zhejiang Lab (second batch of computing satellites launched Oct 14, 2026), LandSpace, state agenciesNational champions, protected by their governmentsAnglo-Persian: the state buys in because fuel is strategicChinese launch cadence; U.S. export-control rules on space compute

The scenario most worth dwelling on is the second, because it maps so cleanly onto the original story. In the 1870s the scarce thing was not oil but the route to market, and the railroads that owned the routes could set terms until someone built a pipeline. If orbit works, the scarce route is mass to orbit. Google’s prototype riding to space on a SpaceX rocket is a small, vivid preview: a rival refiner paying the railroad’s toll. And SpaceX is already renting terrestrial compute to Google and Anthropic, two competitors of its own xAI. It is, in Part One’s language, Vanderbilt with a refinery, now planning to own the track to orbit as well.

The third scenario follows from the second as surely as the Interstate Commerce Act followed the railroads. A single private company owning most of humanity’s access to orbit, and wanting to fill it with a million satellites, is the kind of concentration that produces a commission. The tools already exist: spectrum licensing, debris rules, launch licensing, export controls, and treaty obligations. Watch for regulators to treat launch the way they eventually treated rail: as a common carrier that must serve all comers at published rates.

How long would the advantage last?

The sharpest version of the hypothesis is about duration: that a space-compute lead would be more like Nvidia’s position than like the lab race. To test it, separate two things that are easy to blur: the launch advantage and the orbital-compute advantage.

The launch advantage looks durable, plausibly more durable than Nvidia’s. Models leapfrog each other in months because they are software, copied and distilled. Rockets are not. SpaceX has landed boosters routinely for years, flies weekly or more, and still carries the large majority of the world’s mass to orbit. Blue Origin achieved its first booster reuse only in April 2026. China, by Musk’s estimate, has about 6 percent of mass to orbit. Reusability compounds: every flight lowers cost, which funds more flights. On the evidence, a rival matching SpaceX’s cost per kilogram is a matter of years, not quarters.

The orbital-compute advantage is much weaker. The chips are Nvidia’s, bought on the same terms as anyone’s (SpaceX’s own filing says orbit doesn’t make processors cheaper). The software and models are the same contest as on the ground. The customer can choose between orbit and the ground every time it signs a contract. A lead in launch translates into a lead in orbital compute only if orbit is actually cheaper. If it is, the launch owner captures a toll, and regulators will eventually come for it. If it isn’t, there is no orbital-compute business to defend.

And if it never gets cheaper than the ground? Then SpaceX is still a big landlord on Earth. That is the underappreciated hedge. A $7.56 billion quarter of compute rental, earned on the ground, means SpaceX doesn’t need orbit to work to be a railroad. Orbit is an option on top of a neocloud business that already exists. The risk is the familiar one from 1873: a terrestrial neocloud is a capital-hungry, debt-funded landlord, and its tenants hold termination rights (Anthropic’s contract, per the S-1, can be ended on 90 days’ notice).

ANALYSISVerdict: PARTLY HOLDS Taking orbital compute seriously is right; Part One undersold it. Launch-as-railroad is the strongest part of the hypothesis, and SpaceX’s launch lead is likely to outlast most AI leads. But durability belongs to the launch business, not automatically to space compute, whose economics still hinge on cutting launch cost roughly tenfold and on unserviced chips paying back. The most likely near-term outcome is scenarios 2 and 4 together: a toll road to a niche. If orbit does become cheap, scenario 3 follows: the Railroad Commission comes to orbit.

CHAPTER 2Are robots the new cars?

Part One ended on the gasoline moment: the car turned a waste fraction of the barrel into the most valuable product in it, and the biggest fortunes of the next era went to people who used cheap fuel to do something new. The obvious question is what plays the car in AI. The hypothesis worth testing: robots are the new cars. Robotaxis first, then humanoid and industrial robots. On this view the path is underweighted by everyone outside a small circle of technologists. Tesla and SpaceX are best positioned because they build physical things at scale. Google, through Waymo, has the technology but is less commercialized and less integrated.

The first half of that holds up well. The second half mostly doesn’t, at least not yet.

The commercial leader is Waymo

Start with the numbers, because they are not close. In March 2026 Waymo said it was giving more than 500,000 paid rides a week, double the 250,000 it reported in April 2025. It completed about 15 million rides in 2025, more than triple 2024. As of late October 2026 it operated in 15 U.S. metros, from Phoenix and San Francisco to Atlanta, Miami, Nashville and San Antonio. It served six airports and ran a fleet of roughly 4,000 vehicles. In February it raised $16 billion at a $126 billion valuation. It has announced Madrid and Toronto and a partnership with Toyota for Japan, and it is testing in London and Tokyo.

Its safety case is also the most developed. Waymo’s own data, covering more than 270 million fully autonomous miles through June 2026, claims 82 percent fewer injury-causing crashes and 95 percent fewer serious-injury-or-worse crashes than human drivers. An independent analysis by the Insurance Institute for Highway Safety found about 68 percent fewer police-reportable crashes per mile across its main cities. These are not small effects.

There are real blemishes. In May 2026 Waymo paused freeway service in several cities after a series of incidents in construction zones and issued a software recall covering 3,871 vehicles. Its stated goal of a million rides a week by the end of 2026 looks out of reach: no newer weekly figure has been published since March. Alphabet doesn’t break out Waymo’s numbers, but the Other Bets segment that houses it lost about $2.05 billion in the third quarter on $372 million of revenue. One outside estimate put Waymo’s revenue run rate at around $355 million in February. This is a business still buying its future.

Still, “less commercialized” is the wrong description. Waymo is the only Western company running a driverless ride business at real scale, with paying customers, in more than a dozen cities. If AI has a Model T moment in transportation, it has started with Waymo.

ANALYSISThe part of the hypothesis that survives is integration. Waymo doesn’t build cars. It fits sensors to vehicles made by Jaguar and others; one estimate put a fifth-generation Waymo at about $75,000 of car plus roughly $120,000 of lidar and other equipment. That is a refiner buying its barrels at retail. Morgan Stanley estimates Waymo’s operating cost at about $1.36 a mile against $0.74 for Tesla. Being commercially ahead and being structurally cheapest are different things, and the oil story says the second wins in the long run.

Tesla’s bet: the factory as the moat

Tesla is making the opposite bet. Rather than perfect autonomy in expensive cars and then cut costs, it is trying to get autonomy working in cheap cars it already builds by the hundreds of thousands. In the third quarter of 2026 Tesla delivered 486,532 vehicles and reported $28.0 billion of revenue, roughly flat year on year. Automotive revenue fell 3 percent and GAAP net income dropped 43 percent to $787 million. Its car business is not growing. What investors are paying for is the next thing.

That next thing is real but small. Tesla says its robotaxi service had logged nearly two million cumulative unsupervised miles by the end of the quarter. Cybercab rides began in a limited part of Austin in early September. On the earnings call, Musk said the Austin Cybercab fleet had grown tenfold in about a month. Texas registration records counted 739 Tesla robotaxi vehicles in early October, and the service runs in several Texas and Florida cities. Some cars there are unsupervised and some still carry monitors; Tesla doesn’t disclose the mix, fleet size or paid rides. Federal safety regulators opened an inquiry in September. Tesla has 1.74 million paying Full Self-Driving subscribers, up 67 percent in a year, which is a meaningful base of data and revenue.

The cost goal is aggressive. Tesla has talked about $0.30 per robotaxi mile. Asked on the call whether it was there, the CFO was blunt: “are we at our goal of $0.30 today? No.”

Then there is Optimus. Musk said in October that Tesla had begun early production on a high-volume line in Fremont, with ambitions of 1 million units a year there and 10 million at Giga Texas. Outside deliveries are not expected until around the end of 2027, and no unit count has been disclosed. The 2025 pay package Tesla shareholders approved includes milestones of 1 million robotaxis in commercial operation and 1 million Optimus robots delivered. Today the robotaxi fleet is in the hundreds to low thousands and Optimus deliveries to outside customers are zero.

ANALYSISTesla is best positioned on cost and manufacturing, which is what eventually decided the car industry. It is not best positioned on proven driverless service, which is what decides the next three years. On the evidence, the hypothesis that Tesla is better placed than Google is a bet on the long run, not a description of today.

China: cheap supply, fragile permission

The Chinese players add two lessons the Western debate misses.

The first is cost. Pony.ai says its seventh-generation autonomy kit costs 70 percent less than the previous one, and it expects a complete 2027 robotaxi, vehicle plus kit, to cost under RMB 230,000, about $32,000. It claims its newest robotaxis have reached per-vehicle breakeven in Guangzhou. It had 1,975 robotaxis at the end of June and was aiming for more than 3,500 by year end, plus an Uber contract for more than 2,000 in Europe. WeRide had more than 1,800 robotaxis across 13 countries, including fully driverless service in Abu Dhabi and Dubai. These are still small businesses: Pony’s second-quarter revenue was $36.2 million and it lost $59.8 million. But on vehicle cost, the gap between a roughly $32,000 Chinese robotaxi and a roughly $195,000 Waymo is about six times. That compares estimates from two different sources and years, so treat it as an order of magnitude. That is the Spindletop pattern from Part One: cheaper supply from new owners.

The second lesson is about permission. Baidu’s Apollo Go was the world’s volume leader, with 3.2 million fully driverless rides in the first quarter of 2026. On March 31, multiple Apollo Go cars stalled in Wuhan. Bloomberg reported that Chinese authorities then suspended new robotaxi permits: no new vehicles, no new cities. Second-quarter rides fell to about 1 million, down roughly 55 percent from a year earlier. One bad day and a regulator shut in the field, exactly as the Texas Railroad Commission did with wells in the 1930s.

On humanoids, China dominates the volume. By one industry estimate, Chinese makers accounted for more than 97 percent of the roughly 19,100 humanoid robots shipped worldwide in the first half of 2026, led by AgiBot and Unitree. Unitree reported 2025 revenue of 1.7 billion yuan, about half of it from humanoids, and more than 5,500 humanoids shipped. When it listed in Shanghai on August 19, 2026, its shares opened 629 percent above the IPO price. Investors in China are not underweighting robots.

And China holds a chokepoint the rest of the industry depends on. Robot joints and EV motors need rare-earth magnets. China produced about 69 percent of the world’s rare-earth oxide in 2025 and does up to 90 percent of the processing. Its broad October 2025 export controls were paused for a year. That pause is due to expire on November 10, 2026, and no extension had been confirmed at the time of writing. The narrower April 2025 controls remain in force. In April 2025, Musk said the controls had hit Optimus production. If AI has a Persian Gulf for robots, it isn’t Taiwan. It’s the magnet supply chain.

The West’s humanoid makers

Western humanoid companies are valued like the future and produce like the present. Figure AI was last valued at $39 billion (September 2025) and has robots working at BMW’s Spartanburg plant; it says its BotQ factory can initially build 12,000 a year. Apptronik raised about $520 million in February 2026 at more than $5.5 billion. 1X takes $20,000 pre-orders for its NEO home robot, though home deliveries could not be confirmed by the end of October. Boston Dynamics, owned by Hyundai, started producing its new electric Atlas in January 2026 and said all of that year’s deployments were already committed, to Hyundai and Google DeepMind. Hyundai is building a U.S. plant it says will make 30,000 robots a year.

Amazon is the quiet giant here. It deployed its millionth warehouse robot in July 2025, across more than 300 facilities. Internal documents reported by the New York Times in October 2025 suggested automation could let it avoid hiring more than 600,000 U.S. workers by 2033. Amazon said those documents do not reflect its whole strategy. Its Zoox robotaxi began charging fares in Las Vegas in August 2026 under a federal exemption for up to 2,500 vehicles a year. If robots are the new cars, Amazon is already the biggest fleet operator nobody counts.

The economics: gas stations, not just engines

Why might robotaxis be AI’s gasoline moment? Because the market they attack is enormous and their unit economics are improving fast. By IBISWorld’s estimate, global auto manufacturing is about a $2.6 trillion industry in 2026. AAA puts the cost of owning a new car at $11,577 a year, roughly 77 cents a mile at 15,000 miles. Americans burned about 8.9 million barrels of gasoline a day in 2025, according to the EIA, a little below 2019. A technology that can move a person for less than the cost of owning a car doesn’t compete with taxis. It competes with car ownership itself.

The price gap is closing. A survey of San Francisco fares in late 2025 found an average Waymo ride cost $19.69, against $17.47 for Uber and $15.47 for Lyft, and the premium was narrowing. Forecasts for the robotaxi market are almost useless for precision: ARK talks about a multi-trillion-dollar ecosystem by 2030, while one market-research firm projects about $46 billion. The hundredfold difference is mostly about definitions. The direction is not in dispute.

The oil story suggests where value will collect. Three layers matter:

  • The engine (autonomy software and chips). Waymo, Tesla, Baidu, Pony, plus Nvidia selling platforms to everyone else. Today this looks like refining in the 1860s: several credible players and rapidly falling costs.
  • The car (vehicle manufacturing). Whoever can build cheap, durable, sensor-laden vehicles by the hundreds of thousands. This is where Tesla and the Chinese automakers are strongest, and where Waymo depends on partners.
  • The station (demand and fleet operations). Ride-hail networks, charging, cleaning, depots, insurance. Uber is positioning as the aggregator for everyone’s robots. It has Waymo in Austin and Atlanta, Baidu in Dubai, Pony and WeRide in Europe, and a plan with Nvidia for 100,000 vehicles. That is the branded service station that sells every refiner’s fuel.

In oil, the car created Ford, then General Motors, then a world of suppliers, dealers and stations. None of them were refiners. If the analogy holds, the largest pools of robot profit will go to whoever combines a cheap vehicle with a high-utilization network, not to whoever has the best demo.

ROBOTAXIS AND ROBOTS: WHO’S WINNING, OCTOBER 2026Figures from company filings and reports; Tesla, Waymo fleet data partly from trackers; analysis in the last column
PlayerWhere it standsStrengthWeaknessOil-era analogue
Waymo (Alphabet)500,000+ paid rides a week (March 2026); 15 U.S. metros; ~4,000 vehicles; six airports; $126B valuation (Feb 2026)Years of driverless miles; safety data; the only Western service at real scaleExpensive vehicles built by others; slow fleet growth; a May 2026 freeway pause and recallThe first refiner with a real product. Commercially ahead, not yet integrated.
TeslaRobotaxi in Austin and several Texas/Florida cities; Cybercab rides since September 2026; fleet and paid rides undisclosed; 1.74M FSD subscriptionsBuilds its own cars at scale; low cost per mile (Morgan Stanley est. $0.74 vs Waymo $1.36); a huge camera-equipped fleetStill small as a service; regulators asking questions; Optimus has no disclosed outputFord’s position: owns the factory. Needs the service to work.
Baidu Apollo Go~1M driverless rides in Q2 2026, down from 3.2M in Q1 after a Wuhan outage and a permit freeze; 28 cities; Dubai service via UberScale in China; overseas pushA single outage froze its growth; Baidu’s core revenue shrinkingA big field that a regulator can shut in overnight.
Pony.ai / WeRidePony: 1,975 robotaxis (June 2026), target 3,500+; WeRide: 1,800+ robotaxis, 13 countriesCheap Chinese vehicles (Pony targets < RMB 230k, ~$32k, all-in for 2027); Uber deals in Europe and the Middle EastSmall revenue (Pony $36M in Q2); losses; geopoliticsSpindletop-style low-cost supply looking for export markets.
Zoox (Amazon)Paid rides in Las Vegas since August 2026; purpose-built vehicle; NHTSA exemption for up to 2,500 a yearAmazon’s balance sheet; no steering wheel by designOne paid city; tiny fleetA well-funded latecomer refinery.
Uber / LyftWaymo exclusive on Uber in Austin and Atlanta; Uber–Nvidia plan for 100,000 Level 4 vehicles (first in 2027); Pony, WeRide, Baidu partnershipsDemand aggregation across many robot fleetsOwns no autonomy; robotaxi makers can go directThe branded service station: sells everyone’s fuel.
NvidiaDRIVE Hyperion platform, Cosmos and GR00T models; robotics now inside a $7.2B “Edge Computing” segmentSells to every contenderTesla and Waymo use their own stacksThe rig and barrel maker, again.
Unitree / AgiBot (humanoids)Chinese makers shipped >97% of ~19,100 humanoids in H1 2026 (estimate); Unitree opened 629% above its IPO priceCost and volume; a domestic component baseTiny absolute numbers; mostly research, education and pilotsThe 1890s horseless-carriage shops: many makers, few customers yet.

ANALYSISVerdict on the hypothesis. “Robots are the new cars” is plausible and probably the best candidate for AI’s gasoline moment. “Underweighted” is half right: public attention lags, but capital markets are not ignoring it (Waymo at $126 billion, Unitree’s debut, a Tesla pay package built around robots). “Tesla and SpaceX are best positioned” is a reasonable long-run bet on manufacturing cost, but the evidence today says Waymo is commercially ahead by a wide margin and Chinese makers lead on robot volume and vehicle cost. The structural weakness to watch at Waymo is not commercialization. It is that it doesn’t build its own cars.

CHAPTER 3The survivor and the incumbent

Two of the biggest companies in AI got short shrift in Part One. Meta barely appeared. Google was cast as Standard Oil on structure alone. This chapter tests two hypotheses about them. The first: Meta is underweighted, and Zuckerberg, the one founder who has steered a company through every platform shift of the past two decades, belongs near the center of the story. The second: Google is Standard Oil, but Standard after its prime, and OpenAI is the company best placed to take the AI advertising market away from it.

Meta: two decades of platform shifts

Start with the track record, because it is the strongest part of the case. When Facebook filed to go public in 2012, its prospectus contained a sentence the Acquired podcast’s David Rosenthal calls “the kicker”: “We do not currently directly generate any meaningful revenue from the use of Facebook mobile products.” Ads lived in the right-hand column of the desktop website. As Rosenthal put it, “There is no right hand column on a mobile app.”

The company had bet on mobile web technology rather than native apps. Zuckerberg was blunt about it in a 2024 interview with Acquired’s Ben Gilbert and David Rosenthal: “Well, we were wrong. It turned out that you know, having the native integration was actually critical for having the interactions feel good.” Facebook rebuilt its apps and put ads into the feed. By the fourth quarter of 2013, mobile advertising was about 53 percent of ad revenue, according to its earnings release. It bought Instagram for $1 billion in 2012 and WhatsApp for $19 billion in 2014. Both were derided at the time and are now among the best acquisitions in corporate history. When Apple’s 2021 App Tracking Transparency change cut off much of the data its ad targeting relied on, Meta rebuilt that system too, this time with machine learning.

The record isn’t spotless. The Portal video-calling device was discontinued. The Libra/Diem stablecoin was abandoned and its assets sold for $182 million in 2022 after regulators made clear it could not proceed. And the metaverse has been the most expensive bet in the history of consumer technology. Reality Labs has lost about $96 billion since 2020, computed from Meta’s annual reports through the third quarter of 2026. Meta guides to an $18 to $19 billion loss this year and about $15 billion next, as it moves spending “toward other priority areas, primarily AI.”

What carries Meta through its mistakes is control and scale. Acquired’s hosts make the point repeatedly: Zuckerberg controls the board, so he can absorb losses that would get another CEO fired. On the company’s moat, Gilbert put it in four words: “Everything at this company is scale economies.” Zuckerberg himself, describing the 2022 decision to keep spending on AI infrastructure and Reality Labs, said: “I knew that a lot of the investors would hate it, at least in the short term before it’s clearly the right thing to do.”

The AI-for-ads flywheel

Meta is the cleanest example in the industry of AI paying for itself, because it consumes most of its own output. Its models don’t need to win a chatbot war to make money. They rank feeds, choose ads, write ad variations and measure results, all inside a business with 3.64 billion daily users.

The numbers are strong. Third-quarter 2026 revenue was $64.6 billion, up 26 percent. Ad impressions grew 10 percent and the average price per ad 14 percent: advertisers paid more per ad because the ads worked better. On the second-quarter call, Meta reportedly said its automated Advantage+ ad suite had reached a $75 billion annual run rate, and that its newer ranking models had lifted ad conversions on Facebook by about 16 percent. The CFO said in October that “our AI investments are already delivering substantial incremental advertising revenue.”

Meta also owns its infrastructure to a degree only the hyperscalers match. Gilbert again: “They operate a hyperscaler-sized data center footprint, give or take.” Its 2026 capex guidance is $135 to $145 billion. The Hyperion campus in Louisiana, financed in a joint venture with Blue Owl, is being scaled to five gigawatts and more than $50 billion. And it now has its own frontier models. After the disappointment of Llama 4, Meta Superintelligence Labs released Muse Spark in April 2026 and Muse Wonder, its first frontier model, on October 12, priced at $4 per million input tokens and $20 per million output. Both are closed. In 2024, Zuckerberg had said of open-sourcing: “I’m not like a zealot on this.” He meant it. More than 10 million people in the U.S. and Canada have set up accounts with its Muse personal agent.

The strain is visible, though. Meta’s free cash flow was negative $2.12 billion in the third quarter, its first negative quarter in more than a decade. On October 26 it announced plans to raise up to $95 billion in equity: $35 billion underwritten now, and up to $60 billion over time in new non-voting Class C shares, a structure that preserves Zuckerberg’s control. Nvidia committed $3 billion and AMD $1 billion. Meta also rents capacity it doesn’t own, including roughly $35 billion of contracts with CoreWeave and up to $27 billion with Nebius. And 97.6 percent of its revenue is still advertising. For all its integration, Meta is the least diversified of the big platforms.

Glasses, wristbands, or Neuralink?

Meta is also the only big company still spending seriously on face computers, and the one product line where its hardware is winning. EssilorLuxottica, which makes Ray-Ban Meta and Oakley Meta glasses, sold more than 7 million AI glasses in 2025, against about 2 million in 2023 and 2024 combined. IDC gives Meta about 69 percent of the combined market for headsets and glasses. Meanwhile Apple’s Vision Pro shipped roughly 45,000 units in 2025, and Meta’s own Quest headsets fell 16 percent in the first three quarters of that year. In 2024 Zuckerberg made his conviction plain: “I’m pretty sure glasses and holographic presence and AR is going to be a completely ubiquitous product.”

The momentum is slowing, though. EssilorLuxottica said smart-glasses sales grew about 20 percent in the third quarter of 2026, down from nearly 100 percent the quarter before.

The hypothesis to test is that the real endpoint is not glasses but a direct brain interface: the future is more Neuralink than Ray-Ban. On the evidence, that is a long-run possibility, not a forecast for this decade. Neuralink had implanted roughly 27 people by an unofficial July 2026 count. It raised $650 million in 2025 at about a $9 billion valuation, and its Blindsight vision implant has a breakthrough-device designation but no confirmed human implant. Competitors are also small: Precision Neuroscience has FDA clearance for an array implanted for up to 30 days, and Paradromics did its first human implant in June 2026. Brain implants are measured in dozens of patients. Smart glasses are measured in millions of units a year. Seven million against a few dozen is not a race yet.

The most interesting product sits between the two. Meta’s $799 Ray-Ban Display glasses ship with the Meta Neural Band, a wristband that reads electrical signals from the muscles in the forearm (electromyography) to control the glasses with tiny finger movements. That is a neural interface without surgery. If the long-run future does run through the nervous system, Meta has a consumer product on that road today. Neuralink’s path runs through medical need first, which is how most implanted devices reach scale, if they do.

ANALYSISOn hardware, the hypothesis that Meta’s record is weaker than Musk’s holds. Meta’s big hardware bets (Portal, Quest, a roughly $96 billion Reality Labs bill) have mostly disappointed, and its one clear hit depends on a partner, EssilorLuxottica, to design and manufacture it. Musk’s companies deliver hundreds of thousands of cars a quarter and fly most of the world’s rockets. But on interfaces specifically, Meta is ahead of Neuralink in the only metric that matters for a consumer platform: people using it.

Google: Standard Oil after its prime?

The Standard Oil of 1911 was not a weak company. It was still the largest oil company in the world. But its share of refining had slipped from 88 percent in 1890 to about 64 percent, the market had moved to Texas and to gasoline, and the government was about to break it up. As Acquired’s Ben Gilbert put it in their Standard Oil series, “paradigms break monopolies.” The hypothesis is that Google is in that position: still dominant, structurally the most complete AI company, but past its peak and exposed to a new paradigm it didn’t lead commercially.

Acquired framed it the same way in its 2025 episode on Google and AI: “this is Google today and in perhaps the most classic textbook case of The Innovator’s Dilemma ever.” Google invented the transformer, then watched OpenAI turn it into the consumer product of the decade. And the economics of AI answers threaten the economics of search: “Despite more value creation, there’s way less value capture. Google makes something like $400-ish per user per year just based on some napkin math in the US.”

The evidence for “past peak” is real but thin so far. Google Search & other revenue grew 12 percent in the third quarter of 2026, to $63.35 billion. That is a sharp slowdown from 19 percent in the first quarter and 17 percent in the second, and roughly flat on the quarter before. Management blamed currency moves first, along with lower monetization in AI Mode, and warned of a few more points of pressure in the fourth quarter. Shares fell about 8 percent after the report. Alphabet’s free cash flow was about negative $9.1 billion in the quarter, and it raised money in both the bond and equity markets. Courts have found Google an illegal monopolist twice. The search remedies (no Chrome divestiture, but limits on exclusive deals and data-sharing obligations) are on appeal. In September 2026 the ad-tech court rejected a breakup and imposed six years of behavioral rules.

But the counterattack is also real, and on most measures it is working:

  • Distribution. AI Overviews had more than 2.5 billion monthly users and AI Mode more than 1 billion as of May 2026. The Gemini app reached 1 billion users in October. Google still had about 90 percent of global search share in September 2026 (StatCounter). Management says search query growth is the fastest in five years. Ads are scaling inside AI Overviews and AI Mode, where queries are three times longer than traditional searches.
  • The pipelines. Google is the one company that has, in Rosenthal’s words, both “a foundational frontier model” and “an AI chip.” Nearly 100 customers use TPUs. Anthropic has access to up to a million and, per a Broadcom filing, about 3.5 gigawatts more from 2027. TPU system sales were a “meaningful contributor” to Google Cloud’s 103 percent growth to $30.7 billion in the quarter, and the cloud backlog stood at $561 billion.
  • The models. Google says its new Gemini 4 Carbon leads the Artificial Analysis intelligence index, a company claim. Apple chose Gemini as the basis for its next-generation foundation models in January 2026.
  • The scale. Total revenue grew 25 percent to $127.6 billion. Pichai described a “half-trillion dollar revenue run rate.”

Is OpenAI best placed to win AI advertising?

OpenAI has the strongest consumer brand in AI and more than a billion weekly users. It started testing ads on February 9, 2026, for users on its free and $8-a-month Go tiers, at the bottom of answers and away from sensitive topics. It says it passed a $100 million annualized ad run rate within about six weeks and a $1 billion run rate around September 1, when it opened self-serve ad buying in India, Europe, the Middle East and North Africa. Internal projections reported by Axios call for $2.5 billion of ad revenue in 2026, $11 billion in 2027, $25 billion in 2028, $53 billion in 2029 and $100 billion in 2030, assuming 2.75 billion weekly users by then.

Put that next to the incumbents and the picture is clear.

THE AI ADVERTISING RACE, Q3 2026Quarterly revenue; OpenAI figure is a company run-rate claim, divided by four for scale
BusinessQ3 2026 revenueGrowth (y/y)AI angleSource
Google Search & other$63.35B+12% (down from +19% in Q1)AI Overviews 2.5B+ monthly users; AI Mode 1B+; ads now scaling inside bothAlphabet Q3 2026 release; I/O 2026
Meta advertising$63.06B+26%Impressions +10%, price per ad +14%; Advantage+ reported at a $75B run rate in Q2Meta Q3 2026 release; Q2 call summary
Amazon advertising$19.64B+11% (Prime Day moved to Q2)Sponsored prompts inside its Rufus/Alexa shopping agentAmazon Q3 2026 release
YouTube ads$11.21B+9%—Alphabet Q3 2026 release
OpenAI (ChatGPT ads)~$0.25B a quarter (company-stated ~$1B annualized run rate, Sept 2026)Launched Feb 2026Ads on Free and Go tiers only; internal target $2.5B for 2026OpenAI via TechAfrica News; Axios via Reuters

OpenAI’s ad business, at a claimed $1 billion annualized, is under half of one percent of Google’s search revenue. Its own forecast would get it to roughly where Amazon’s ad business is today by around 2029. The firm eMarketer expects all standalone U.S. chatbots combined to earn less than $1 billion of U.S. ad revenue in 2026. Meanwhile OpenAI’s commerce push has retreated. It moved checkout out of ChatGPT in March 2026; Shopify’s president said only about a dozen of its merchants were selling through AI tools. It shut its Atlas browser in August, saying AI browsers “should be a feature, not a destination.” Its center of gravity is shifting too: the CFO reportedly told investors in August that enterprise revenue now exceeds consumer.

None of that means the threat is fake. Ahrefs estimated in early 2026 that search-like prompts on ChatGPT were already about 12 percent of Google’s search volume, using mid-2025 data, while Google still sent about 190 times more traffic to websites. The Acquired hosts identified the real risk: “AI might take away the majority of the use cases of Search, and even if it doesn’t I bet it takes away a lot of the highest value ones.” Losing the most valuable commercial queries at the margin would hurt Google’s revenue far more than the volume share suggests.

ANALYSISVerdicts. PARTLY HOLDS on Google as Standard after its prime: the deceleration in search, weaker monetization in AI Mode, negative free cash flow and two antitrust losses fit, but this is a company growing 25 percent with its own chips, models and the world’s largest distribution. It looks less like Standard in 1911 than like Standard around 1901, the year of Spindletop: still dominant, facing new supply from new owners, with the decline optional rather than fated. DOESN’T HOLD (YET) on OpenAI as best placed to dominate AI ads. It has the users and the intent signals, but Google and Meta already run ads inside AI answers at a scale hundreds of times larger, Meta’s ad system is arguably the most AI-native in the world, and OpenAI’s own revenue is tilting toward enterprise. The likelier outcome is that OpenAI becomes a meaningful new ad platform while Google’s share of the highest-value queries erodes at the edges. That is a 1901 outcome, not a 1911 one.

ANALYSISOn Meta: HOLDS on underweighting. Meta owns infrastructure, has closed frontier models, a captive use for them, and the best AI-for-ads results in the industry. In oil terms it is an integrated major that burns most of its own refined product in its own engine. The caveats are concentration (ads are nearly all its revenue), a balance sheet now leaning on outside capital, and a model lab that only reached the frontier this month.

CHAPTER 4Rebates, recast

Part One cast the hyperscalers as railroads and the labs as refiners, with Nvidia as Vanderbilt collecting the toll. Here is a sharper version of that map, offered as an objection. If the hyperscalers and neoclouds are the railroads, then what they want, like every railroad, is guaranteed volume. The companies that control volume, the big AI labs, are therefore Standard Oil. Rockefeller got his rebates from Vanderbilt because he could promise sixty carloads a day. So the labs with the most volume, Anthropic or Google, will get better rates from SpaceX and everyone else than small buyers can. The rest of the hypothesis follows. Someone will consolidate the fragmented neoclouds, the way Standard swallowed 22 of 26 Cleveland refiners in six weeks. And Oracle is the Pennsylvania Railroad overplaying its hand.

It is a better map than it first sounds. Here is how it holds up.

Volume and rebates, 2026 edition

The first part is clearly right: the whole industry runs on volume guarantees. Rockefeller’s Lake Shore deal was a promise of steady freight in exchange for a better rate. The 2026 equivalent is the take-or-pay contract, and it now runs through every layer of the chain.

THE TAKE-OR-PAY CHAINSelected commitments; figures from filings and reports as of Oct 2026
Who promises whomWhat’s been committedSource
Memory makers → chip and cloud buyersMicron’s strategic customer agreements are “structured as take-or-pay,” with price bands and cash deposits; SK hynix has long-term agreements with 14 customers covering more than half its 2027–28 capacity, with advance payments of roughly 10–30%Micron FY2026 10-K; TrendForce; SK hynix Q3 results
Nvidia → its suppliersSupply and capacity commitments of $279B as of July 26, 2026, up from $119B a quarter earlierNvidia 10-Q
Nvidia → neocloudsObligated to buy CoreWeave’s unsold capacity through April 2032 ($6.3B initial order); leasing back 18,000 GPUs from Lambda for $1.5B; guarantees of up to $105B on Ohio data-center leases for an OpenAI affiliateCoreWeave 8-K; Tom’s Hardware; Nvidia 10-Q
Hyperscalers → neocloudsMicrosoft–Nebius $17.4–19.4B; Microsoft–IREN ~$9.7B with a 20% prepayment; Google backstopping Fluidstack leases at TeraWulf and Cipher sitesNebius filing; IREN; TeraWulf release
Labs and platforms → landlordsAnthropic: $265B+ of non-cancelable commitments in its S-1; $1.25B a month to SpaceX through May 2029; $35B to Lambda. Meta: ~$35B with CoreWeave, up to $27B with Nebius. OpenAI: ~$22.4B with CoreWeave; ~half of Oracle’s $664B backlogAnthropic S-1; SpaceX filing; TechCrunch; CoreWeave; ERP Today

Do volume buyers get rebates? Yes, in three forms.

Price. Commitment discounts are the posted version of the rebate. Google Cloud’s listed H100 node price falls about 31 percent with a one-year commitment and 56 percent with a three-year one; Azure’s reservations run 25 to 56 percent off. Hyperscaler on-demand list prices for an H100 hour run $7 to $12, against roughly $4 to $6 at the neoclouds and about $3 on a market index (Ornn’s settled index was $3.03 on October 30). The biggest buyers negotiate privately below all of these.

Priority. In a shortage, the scarce thing isn’t a discount but a place in line. Here the evidence cuts the other way. Anthropic’s deal with SpaceX, $1.25 billion a month for roughly 300 megawatts and more than 220,000 Nvidia GPUs, works out to about $50 million per megawatt-year. That is roughly $7.80 per GPU-hour, around two and a half times the H100 market index. The comparison is loose, because the fleet includes newer chips and the price includes power and facilities. But it doesn’t look like a rebate. It looks like a big buyer paying up for capacity that exists now. In May, Anthropic said the deal let it double Claude Code’s rate limits, which tells you how badly it needed the space. Rockefeller rarely had to pay for urgency. The labs of 2026 do.

Credit. This is the modern rebate people miss. A contract with an investment-grade tenant is collateral. CoreWeave’s $8.5 billion facility in March 2026, anchored by Blackstone, was the first investment-grade financing secured by GPU infrastructure “and an associated customer contract,” priced around SOFR plus 2.25 percent. Nebius’s first secured facility, backed by an investment-grade customer, priced at SOFR plus 2.5 percent. CoreWeave’s unsecured notes, by contrast, carry a 9.75 percent coupon. A Microsoft, Meta or Google contract lets the railroad borrow cheaply, and some of that saving flows back to the tenant in price. Volume buys cheaper capital, which buys cheaper capacity.

And SpaceX renting to both Anthropic and Google is the hypothesis’s best single piece of evidence. A company with its own frontier lab is selling capacity to two direct competitors on large, multi-year commitments, because its rails need freight. That is the Lake Shore and Flagler, almost exactly.

The neocloud roll-up that hasn’t happened (yet)

If the neoclouds are the 1860s refiners of Cleveland, numerous, leveraged and dependent on a few big customers, then consolidation should be coming. The stress is visible:

  • CoreWeave had about $104 billion of backlog at the end of June and second-quarter revenue of $2.58 billion, up 112 percent. But net interest expense was $640 million, about a quarter of revenue, on roughly $35 billion of debt. Its five-year credit default swaps blew past 1,100 basis points in mid-October, implying roughly a 64 percent chance of default over five years by one market calculation. The stock closed at $54.87 on October 30, about 60 percent below its September 2025 level.
  • Nscale’s IPO filing showed first-half 2026 revenue of $140.6 million, a net loss of $1.02 billion, and its biggest customer at 52 percent of revenue, against $103 billion of contracts, most not yet active.
  • Lambda pushed its IPO to 2027 and is raising up to $4 billion. Its backlog jumped from $15 billion to $50 billion in a quarter, mostly on a $35 billion Anthropic commitment.
  • Fluidstack was valued above $18 billion with more than $15 billion of debt. Crusoe raised at $30.9 billion and Nebius reported $582 million of quarterly revenue and $3 billion of annualized revenue, with Microsoft and Meta as anchors.

Yet the one attempted roll-up failed. Core Scientific’s shareholders voted down CoreWeave’s all-stock takeover on October 30, 2025. No major neocloud merger has been verified in 2026. Who could play Rockefeller here? The candidates, in order of how much they already look the part:

  1. Nvidia. It has stakes in CoreWeave, Nebius ($2 billion in March 2026), Nscale and Crusoe. It backstops their unsold capacity, rents GPUs back from them, and arranged more than $500 billion of financing capacity with six asset managers. Acquired’s David Rosenthal noted in 2023 that Nvidia “has also been seeding new cloud providers out there in the ecosystem.” Nvidia doesn’t need to buy the neoclouds. It already holds them the way the Standard Oil Trust held its member companies: through ownership stakes and dependence, not one corporate name.
  2. The anchor tenants. Anthropic’s commitments now underwrite a large share of several neoclouds’ backlogs. Microsoft and Meta anchor others. A tenant that is half your backlog effectively decides your fate.
  3. Private credit. Blackstone, Apollo and Blue Owl finance the assets. In Anthropic’s $35 billion chip-lease vehicle with Apollo and Blackstone, Broadcom even guarantees the chips’ resale value if Anthropic defaults. If the neoclouds hit distress, the lenders own the GPUs. That is the J. P. Morgan path of the 1890s, when bankers reorganized bankrupt railroads into a few systems, more than the Rockefeller path.
  4. SpaceX. Already larger than any neocloud by quarterly compute revenue, with access to public equity markets since its June IPO. On October 26 it locked in about $40 billion of GPU lease financing with Apollo, and Musk predicted a wave of death and consolidation among the more than 200 neoclouds.

ANALYSISThe Cleveland Massacre happened under duress: refiners sold because the South Improvement scheme threatened to cut them off from rail. The modern equivalent would be a neocloud funding crisis, and the CDS market is pricing one at the edges. Consolidation is likely. The first round, though, is more likely to come through lenders and anchor tenants taking assets in distress than through one company buying rivals in good times.

Oracle and the Empire Transportation Company

In 1877, the Pennsylvania Railroad’s Empire Transportation subsidiary moved into refining, the business of its biggest customer. Standard pulled its freight, the railroad blinked, and Standard bought Empire’s assets for $3.4 million. The hypothesis casts Oracle as the Pennsylvania Railroad overplaying its hand.

On overextension, the evidence is strong. Oracle’s remaining performance obligations reached $664 billion at the end of August 2026, and roughly half of that is tied to OpenAI. Its total debt was $125 billion. Free cash flow was negative $5.4 billion in the quarter, and capex guidance for the fiscal year is $90 to $95 billion. S&P cut Oracle to BBB-, one notch above junk, in July 2026. It projected a free operating cash flow deficit of nearly $42 billion for the fiscal year and cited “extreme customer concentration risk.” Oracle’s credit default swaps have risen to around 300 basis points. Its stock closed at $130.67 on October 30, about 62 percent below its September 2025 record of $345.69. When OpenAI told investors in October that its annualized revenue was about $50 billion, not the widely reported $68 billion, Oracle fell about 6 percent in a day.

On the specific Empire parallel, the fit is weaker. Empire’s sin was moving into its customer’s business. Oracle isn’t building a frontier model to compete with OpenAI. Its exposure is the older railroad sin: building ahead of demand, on borrowed money, for one big shipper who has alternatives. OpenAI also has capacity deals with Microsoft, AWS ($38 billion), CoreWeave and AMD. That makes Oracle look less like the Pennsylvania Railroad in 1877 and more like Jay Cooke’s Northern Pacific in 1873, the railroad financing that set off a panic. The Empire move, an infrastructure owner going into its customers’ business, is being made today by SpaceX, Google and Meta. Those companies, unlike the Pennsylvania in 1877, are stronger than their tenants.

Three maps, compared

Rockefeller’s ultimate advantage was not rebates. It was that Standard became everything: refinery, pipeline, tank car, barrel works and tank-wagon delivery. On that test, the integration scorecard looks like this:

WHO OWNS WHICH LAYERAnalysis based on company disclosures
CompanyChipsCloud / computeFrontier modelsDistributionRead
GoogleYes (TPU)Yes (Google Cloud)Yes (Gemini)Yes (Search, Android, Chrome, YouTube)Most complete
SpaceX / xAIPlanned (Terafab); buys NvidiaYes (1.7 GW, rents to rivals)Yes (Grok)Partly (X, Cursor, Starlink)Owns launch and connectivity too
MetaPartly (MTIA); buys Nvidia, AMDOwn data centers, no external cloud; also rentsYes (Muse, closed)Yes (3.6B daily users)Integrated around a captive use
MicrosoftPartly (Maia)Yes (Azure)Partly (stake in OpenAI; own models)Yes (Office, Windows, GitHub)Integrated through partnership
AmazonYes (Trainium)Yes (AWS)Partly (stakes in Anthropic, OpenAI; Nova)Yes (retail, Alexa)Integrated through partnership
OpenAIAnnounced custom chip with BroadcomPartly (Stargate sites, rented)YesYes (ChatGPT)Brand and distribution; rents the rails
AnthropicNo (uses TPU, Trainium, Nvidia)No (rents from Google, Amazon, Microsoft, SpaceX, neoclouds)Yes (Claude)Partly (Claude apps, Claude Code)Mostly rents

That points to a third map, also worth testing. In it the hyperscalers are the refiners, consolidating from many to few, and Musk is Rockefeller: the integrator who ends up owning launch, connectivity, compute, models and distribution. Here are the three maps side by side.

THREE MAPS OF AI, SCOREDAnalysis — a judgment call, not a finding
MapCastingsEvidence forEvidence againstBest used for
A. Part One’s mapNvidia = Vanderbilt’s toll; hyperscalers = railroads; labs = refiners; Google = StandardMargins: Nvidia’s 75% gross margin is the toll; open weights compress refiningUnderplays how much leverage volume buyers have, and SpaceX’s pivot to landlordBest guide to where profit sits today
B. The volume mapHyperscalers and neoclouds = railroads craving guaranteed volume; labs and volume controllers = Standard, winning rebates; Oracle = Pennsylvania RailroadTake-or-pay contracts everywhere; commitment discounts of 25–56%; big tenants’ credit lowers landlords’ borrowing costs; SpaceX renting to Anthropic and GoogleThe volume controllers aren’t taking the margin: compute is about 55% of Anthropic’s revenue; suppliers earn the fat marginsBest guide to how contracts and power are being negotiated
C. Elon as RockefellerHyperscalers = refiners consolidating from many to few; Musk = the integrator who becomes everythingSpaceX now earns more from AI than from rockets; bought xAI, X and Cursor; owns launch, connectivity, compute and modelsNot the low-cost or dominant refiner; xAI’s models aren’t the most used; the empire is debt- and equity-financedBest guide to the ambition, not yet the position

The decisive test is where the money ends up. In the 1870s, the volume controller took the margin and the railroads fought rate wars. In 2026, look at the layers’ economics. Nvidia earns a 75 percent gross margin. Google Cloud earned about a 36 percent operating margin last quarter, and AWS about 39 percent. Anthropic’s S-1 shows compute costs of $8.9 billion, 55 percent of first-half revenue. The FT reported it earned about $1.70 of revenue per dollar of compute in the third quarter. CoreWeave lost money after interest. Profit is concentrated at the chip layer and the hyperscalers’ clouds, not at the volume controllers. That is Map A’s world, not Map B’s.

But Map B describes the direction of travel. The labs are behaving like Rockefeller. Anthropic buys compute from Google’s TPUs, Amazon’s Trainium, Nvidia through Microsoft and SpaceX, and a roster of neoclouds, playing railroads off against each other exactly as Flagler did. Its third-quarter numbers suggest the playbook is starting to work. If custom chips and abundant capacity erode Nvidia’s toll, the margin should migrate toward whoever controls the volume and the customer.

ANALYSISVerdict: PARTLY HOLDS Volume-for-rates is the right lens for the contracts. SpaceX renting to rivals is the Lake Shore deal reborn, and consolidation of the neoclouds is likely. But margins say the toll collector (Nvidia) and the integrated landlords still capture more than the volume controllers. Oracle fits “railroad overplaying its hand” but not the Empire specifics; the Empire move is being made by stronger infrastructure owners. The race to become Standard is a race among integrators. Google is most complete; SpaceX/xAI and Meta are closing in from different directions. Anthropic is the most Rockefeller-like in strategy and the least in ownership. Musk as Rockefeller (Map C) is a description of the ambition, not yet of the position.

CHAPTER 5Breadth, cycles, and the moving chain

Revenue breadth as survival

Part One ended its bubble discussion with a distinction. The railroads of the 1870s were standalone companies living on debt, while today’s biggest AI builders have huge, profitable core businesses behind the spending. That pushes the 1873-style risk to the edges: debt-financed neoclouds, vendor-financed customers, and investors paying peak multiples. The hypothesis here extends it. Revenue breadth is survival. The companies with the broadest revenue will outlast single-source players. Pure token production loses value fast, especially against open models, so value flows either to the most valuable tokens or to whoever helps people use them.

REVENUE BREADTH SCORECARDShares of revenue from the latest reported quarter; breadth and fragility columns are analysis
CompanyRevenue mixBreadthMain fragilitySource
AmazonOnline stores 34%, third-party seller services 23%, AWS 23%, ads 10%, subscriptions 7%BroadestNegative free cash flow from the build-out; retail margins thinQ3 2026 release
AlphabetSearch 50%, Cloud 24%, subscriptions & devices 11%, YouTube ads 9%, network 6%BroadSearch is half of revenue and the part AI threatens; negative FCF in Q3Q3 2026 release
MicrosoftAgents & Infra 83% (Azure alone ~36%), Devices & Consumer 17%BroadHeavy exposure to OpenAI; $677B commercial backlog to deliverFY27 Q1 release
AppleiPhone 50%, Services 28%, Mac 9%, wearables 7%, iPad 6%Broad, but little AI revenueDepends on partners (Google’s Gemini) for frontier AIFQ3 2026 release
SpaceX / xAISuper Intelligence 52% (mostly compute rental), Starlink 36%, launch 12%; X ads ~4%Broadening fastAI revenue concentrated in a few tenants with termination rightsQ3 2026 release
MetaAdvertising 97.6%Narrow source, deep demandOne revenue line; now raising outside equityQ3 2026 release
OracleCloud infrastructure 38%, license & support 28%, SaaS 22%MediumAbout half of $664B backlog tied to one customer (OpenAI); BBB-Q1 FY27; S&P
NvidiaData center 92.5%Narrow, but strongestFive direct customers were 10–22% of receivables each; growing guaranteesQ2 FY27 release; 10-Q
BroadcomAI semiconductors ~56% and risingNarrowingA handful of custom-chip customersQ3 FY26 release
MicronMemory; revenue up ~5x in a yearNarrow, cyclicalThe most cyclical layer; cushioned by take-or-pay depositsFQ4 2026 release; 10-K
OpenAIEnterprise now a majority; subscriptions; API; ads ~$1B run rateMedium, single product familyLargest spending commitments relative to revenueReported, Aug–Oct 2026
Anthropic~80% usage-based (API, coding tools), ~20% subscriptions; ~48% through cloud marketplacesNarrowTokens are the product; compute ~55% of revenue (H1)S-1 via Mostly Metrics
CoreWeaveGPU rentalSingle-sourceInterest ≈ 25% of revenue; largest customer 45% (Q1)Q2 2026 release; TIKR

Three things stand out.

Breadth and integration are different things. Meta is deeply integrated but almost entirely dependent on one revenue line, advertising. Apple is broad but has little AI revenue of its own. Amazon is the broadest, and SpaceX has gone from mostly Starlink to mostly compute rental in a year. Breadth protects you from a single market turning. Integration protects your margin within a market. Standard Oil had integration first and breadth (gasoline, lubricants, petrochemicals) later.

The fragile layer is exactly where Part One predicted. CoreWeave’s interest bill ate about a quarter of its revenue last quarter, and its credit spreads price a real chance of default. Oracle has half its backlog tied to one customer and a rating one notch above junk. Nscale’s biggest customer was more than half its revenue. These are single-revenue, debt-financed infrastructure businesses whose customers can choose other railroads. If the cycle turns, this is where it breaks first.

“Pure token production loses value fast” is half right. At a fixed level of capability, prices collapse. Epoch AI estimated in September 2026 that cost at a given performance level has fallen about 47 percent a quarter since 2023. YipitData sees token prices down about 50 percent year over year across its datasets. Open-weight models list at roughly a quarter of closed-model prices, as Part One showed. But the frontier hasn’t deflated nearly as much. GPT-4 launched in 2023 at $30 per million input tokens and $60 per million output. In 2026, OpenAI’s GPT-5.6 Sol lists at $4 and $20, Anthropic’s Opus 5.5 at $4 and $20, and its top Fable 5.1 model at $10 and $50. That is roughly three to seven times cheaper in three and a half years, for vastly more capable models. And Anthropic, whose revenue is about 80 percent usage-based, grew quarterly revenue from $4.7 billion in the first quarter of 2026 to $16.8 billion in the third. Commodity tokens lose value fast. The most valuable tokens, for coding, agents and hard problems, are holding their price and growing in volume. The second escape route is applications that help people use tokens. SpaceX paid an implied $60 billion for Cursor, a coding tool that sits on top of other companies’ models.

ANALYSISVerdict: HOLDS, WITH A CORRECTION Breadth buys time, and the single-source, debt-financed landlords are the most fragile part of the chain. But pure token sellers at the frontier are, for now, among the fastest-growing businesses in the world. Value isn’t leaving tokens. It is concentrating in the most valuable ones, and in the tools that put them to work.

Boom, bust, repeat

Oil’s history is a cycle of gluts and shortages. Titusville crude fell from $20 a barrel to 10 cents in two years; East Texas crashed from $1.15 to 10 cents in 1931; then came the shortages of the 1970s. Economists call this a cobweb cycle. Supply responds to price with a long lag, because wells, refineries and pipelines take years to build. Prices overshoot in both directions.

The hypothesis is that AI runs the same loop. Breakthroughs make models cheaper. Cheap models unlock new demand. Demand pushes the price of the underlying capacity back up. The evidence for this is now strong:

  • GPU rentals. An H100 rented for about $8 an hour at the 2023 peak, fell to a low of about $1.70 in October 2025 as new capacity flooded in, then climbed back. Silicon Data’s index read $2.88 on October 24, 2026, up 44 percent since December; Ornn’s settled index was $3.03 on October 30. In January, Silicon Data expected 2026 prices to “stabilize or decline modestly.” They rose instead, the forecasting miss every cobweb produces.
  • Memory. Conventional DRAM contract prices rose about 93 to 98 percent in a single quarter in early 2026, according to TrendForce. Micron’s quarterly revenue went from $11.3 billion to $54.2 billion in a year, at an 86.8 percent gross margin, higher than Nvidia’s. SK hynix says it has sold more than half its 2027 and 2028 capacity under long-term agreements and expects shortage through 2030. Musk told Dwarkesh in February that his biggest supply concern was memory.
  • Rationing. Anthropic ran reduced peak-hour limits on Claude Code in early 2026. When its SpaceX capacity came online in May, it doubled the limits. Shortage at the bottom of the chain shows up as rationing at the top.

Look at what moved. In 2023 the scarce input was the GPU. By late 2025 GPUs were in glut. In 2026 the scarcity moved to memory and to power (Chapter 1’s turbine backlog, Part One’s Texas queue). This supports the most interesting part of the hypothesis: the value chain doesn’t just ebb and flow, it shifts. Memory makers, a commodity business for decades, are briefly earning higher margins than the chip designer at the center of the boom. In oil terms, the barrel was briefly worth more than the oil again.

Where does business spending settle? It follows usage, because most of it is metered. Anthropic’s revenue is about 80 percent usage-based, and OpenAI’s enterprise revenue overtook consumer in 2026, by its CFO’s reported account. Jevons says “settle” may be the wrong word. In oil, every time a mature use stabilized (lighting), a new one arrived (cars, then aviation, then petrochemicals). Part One’s evidence is that tokens processed are growing far faster than prices are falling. Business AI spending will settle per task. Whether it settles in total depends on whether agents become the car.

Does consumer AI settle at $50 to $100 a month?

One concrete version of “settling” is that consumer AI ends up costing $50 to $100 per person per month. Today’s price menu looks like this:

CONSUMER AI PRICE TIERS, OCTOBER 2026Monthly U.S. prices from company pricing pages and reporting
ProviderTiers (per month)
OpenAI ChatGPTFree (with ads); Go $8 (with ads); Plus $20; Pro $100 and $200
Anthropic ClaudeFree; Pro $20; Max $100; Ultra $200
Google AIAI Pro $19.99; AI Ultra from $99.99 (cut from $249.99 at launch)
xAI GrokSuperGrok Lite $10 to SuperGrok Heavy $300 (per an in-app catalog, Aug 2026)
PerplexityPro $20; Max $200
Amazon Alexa+$19.99 standalone; free with Prime
BenchmarksU.S. streaming: $76/month per subscribing household (Deloitte, fall 2026); home broadband: $81.16/month average (BroadbandNow 2026)

The market has converged on a $20 anchor with $100 and $200 tiers for heavy users, plus a sub-$10, ad-supported entry tier. Google even cut its top plan from $249.99 to $99.99. But very few people pay. OpenAI had 50 million paying subscribers against 900 million weekly users in February 2026, about 5.6 percent. Menlo Ventures’ 2025 survey found only about 3 percent of U.S. consumer AI users paying, in a consumer market it estimated at about $12 billion. Averaged across all users, subscription revenue is a dollar or two a month, nowhere near $50.

But subscriptions aren’t the only way consumers pay. Acquired’s napkin math puts Google’s U.S. ad revenue at “something like $400-ish per user per year,” about $33 a month, paid indirectly through advertising. Add a subscription and commerce fees, and $50 to $100 a month of total monetization per engaged user starts to look plausible over time. For comparison, a U.S. household pays about $76 a month for streaming and $81 for broadband.

ANALYSISVerdict: PLAUSIBLE AS TOTAL MONETIZATION, NOT AS A SUBSCRIPTION PRICE A direct $50–100 monthly subscription for the average person would put AI at the level of a household’s whole broadband or streaming bill, per person. Today only a few percent pay even $20. The more likely path is a blend: a $20-ish anchor for those who pay, ads and commerce for everyone else, and $100–200 tiers for professionals. Together these could reach $50–100 per engaged user, roughly what search and social already earn indirectly from a U.S. user.

Amazon’s shopping test

If agents start doing the shopping, the most exposed and best-placed company is Amazon. The hypothesis says it could win agentic shopping, but only if it gets its act together. On the 2026 evidence, it is getting its act together faster than that suggests, using a strategy straight out of Standard Oil’s tank-wagon playbook: control the last mile and fence it.

Building its own agent. Amazon says its Rufus assistant was used by more than 300 million customers in 2025 and delivered nearly $12 billion of incremental annualized sales (Amazon’s own estimate). In 2026 it merged Rufus and Alexa+ into “Alexa for Shopping.” It says 400 million customers used it over the past year, interactions grew 570 percent, and users spend over 40 percent more per order. It is already selling sponsored prompts inside the assistant, the agent-era version of sponsored products. Its “Buy for Me” feature lets that agent buy from other brands’ websites.

Fencing the store. At the same time, Amazon keeps outside agents out. Its robots.txt file blocks 99 named AI agents, including OpenAI’s, Anthropic’s, Perplexity’s, Google’s, Meta’s and xAI’s. In September 2026 it blocked Meta’s Muse agent under its conditions of use. It sued Perplexity over its Comet browser agent and won a preliminary injunction in March 2026. But the Ninth Circuit vacated it in August, reasoning that the user, not Perplexity, was the one accessing Amazon. The legal fence is weaker than the technical one.

Hedging with the other side. Amazon has also become one of OpenAI’s biggest partners: a $38 billion AWS contract, a $50 billion investment in OpenAI’s February round, and a pilot, announced in September, for Amazon Ads to sell advertising inside ChatGPT. OpenAI, meanwhile, retreated from in-chat checkout after little merchant uptake. McKinsey forecasts that agent-orchestrated retail could reach up to $1 trillion in the U.S. by 2030, a forecast and not data.

ANALYSISVerdict: LEANS AMAZON Amazon has the inventory, logistics, payments, reviews and ad business that an agent needs, and it is monetizing its own agent already. Its vulnerability is that its roughly $20 billion-a-quarter ad business assumes people browse pages; agents don’t look at ads. Its answer is to make its own agent the one people use and sell ads inside it. The risk is legal and behavioral: if courts keep ruling that a user’s agent can shop anywhere, the fence comes down, and Amazon has to win on price and delivery alone. It is well placed to do that.

Nvidia: can the toll collector move across the stack?

Nvidia controls the dominant chip design and on current numbers doesn’t look threatened. Revenue was $96.2 billion in the quarter ended July 26, up 106 percent, with a $108 billion guide. The Acquired podcast’s three-part history explains why the moat has held. The hosts identify scale economies (CUDA’s cost spread over millions of developers) and switching costs as Nvidia’s core powers. Jensen Huang, interviewed on the show in 2023, described the moat as the ecosystem itself: “That network is essentially your moat.”

The question is whether Nvidia can move upstream or across the stack. It has been doing so for years, deliberately and with restraint:

  • From chip to data center. Huang told Acquired that “if you want to be a data center company, building the processing chip isn’t the way to do it,” which is why Nvidia bought Mellanox for networking. Acquired’s Ben Gilbert summarized the thesis: “the data center is the computer.” Rosenthal’s response to Nvidia selling full systems: “you say solution, I hear gross margin.”
  • A cloud, carefully. In 2023 Nvidia launched DGX Cloud, which Rosenthal described as “their own cloud service through other clouds.” The Information reported in September 2025 that Nvidia had scaled it back toward internal use. Nvidia disputed this, calling it “fully utilized and oversubscribed.” Huang’s formulation, quoted on Acquired: “We build our systems full stack, but we go to market in a disaggregated way.”
  • Models. Nvidia publishes its Nemotron family as open models with open weights, training data and recipes, and robotics and driving models (Cosmos, GR00T, Alpamayo). These support its customers rather than compete with the frontier labs.
  • Finance. This is the biggest move. Nvidia put $30 billion into OpenAI’s February 2026 round, after Huang said the earlier $100 billion plan was “probably not in the cards.” It invested $2 billion in Nebius, $5 billion in Intel and $1 billion in Nokia, and committed $3 billion to Meta’s raise. It licensed Groq’s inference technology for $20 billion. It must buy CoreWeave’s unsold capacity through 2032 and has guaranteed up to $105 billion of data-center leases for an OpenAI affiliate. Its 10-Q lists $279 billion of supply commitments and $25 billion of equity commitments to AI model makers. It has also lined up more than $500 billion of financing capacity with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR.

The restraint is the lesson from TSMC. As Rosenthal put it in Acquired’s Morris Chang episode, TSMC is “the only foundry at the leading edge that does not in any way compete with their customers.” If Nvidia launched a full public cloud or a frontier model, it would compete with its biggest buyers, who are already building their own chips. Google has nearly 100 TPU customers. Broadcom’s AI chip revenue grew 221 percent to $16.7 billion last quarter. AMD has 6-gigawatt deals with both OpenAI and Meta. Acquired’s bear case from 2023 still applies: “literally everybody else in the technology ecosystem is now aligned and incentivized to say, I want to take a piece of Nvidia’s pie.”

ANALYSISVerdict: HOLDS, WITH A TWIST Nvidia isn’t threatened in chip design today, and it is moving across the stack: into systems, networking, software and models. But it does so mostly as financier and guarantor, not as a rival cloud or lab. That makes it less like Vanderbilt and more like a railroad that lends its customers the money to buy freight cars. The new risk to Nvidia is not a better chip. It is the balance sheet it has extended to its own customers. If the neocloud layer breaks, the guarantees come home.

CHAPTER 6The players you’re not watching

The AI debate is conducted almost entirely in terms of about ten companies. The oil story warns against that. The biggest fortune of the nineteenth century went to a refiner who started as a produce merchant. The real long-run rents went to sovereigns nobody in Titusville had heard of. And the barrel makers had their moment when the barrel was worth more than the oil. Here are the clusters that matter more than their share of attention, each with its place in the analogy, why it may matter more than people think, and what to watch.

THE NEW BARREL MAKERS

Memory: SK hynix, Samsung, Micron

Where it sits: the barrel in 1861, the container briefly worth more than what it carries.

Why it matters: memory is the most striking shift of 2026. SK hynix’s preliminary third-quarter operating profit was 83.3 trillion won on 104.6 trillion of revenue, an 80 percent margin. Samsung’s preliminary operating profit of 107.4 trillion won in one quarter exceeded its entire 2025. Micron’s gross margin hit 86.8 percent. SK hynix says its HBM and DRAM are sold out through 2029 and expects supply to fall short through 2030. TrendForce estimates DRAM contract prices rose about 80 percent in the second quarter and 40 percent in the third. The memory makers have also adopted take-or-pay contracts with deposits, the clearest sign of who has leverage. The leader is being challenged, though: Counterpoint estimates SK hynix’s share of HBM revenue fell from 64 percent to 50 percent in a year as Samsung rose from 15 to 33.

Watch: whether price ceilings in long-term agreements hold if spot prices keep rising, and TSMC’s warning that rising component prices are already hurting price-sensitive chip demand.

THE NEW RAILROAD COMMISSION

Power, turbines and permission to connect

Where it sits: the Texas Railroad Commission and the landowners together, rationing the input rather than the output.

Why it matters: GE Vernova’s gas backlog and slot reservations reached 122 gigawatts in the third quarter, and its total backlog hit $197 billion. Siemens Energy holds 69 gigawatts with lead times of “three years or more.” Mitsubishi Heavy holds 35 gigawatts, delivering into 2028–2030. Together that is roughly 226 gigawatts reserved against perhaps 25 to 30 gigawatts shipped a year (a rough estimate). Transformer lead times run up to four years. In the U.S. interconnection queue, gas grew 86 percent in 2025 while solar and wind shrank. PJM’s latest capacity auction cleared at its $325-per-megawatt-day cap in every zone and still came up about 6.8 gigawatts short. Without the cap, its simulation put the price at about $555. Price caps are hiding scarcity. Nuclear is moving but slowly: TerraPower got the first construction permit for a commercial non-light-water reactor in March 2026, with completion targeted for 2030, and the U.S. government has a deal for at least $80 billion of Westinghouse reactors. Constellation, now the largest U.S. power producer after buying Calpine, has 20-year contracts with Microsoft.

Watch: the outcome of the Texas audit; FERC’s region-by-region large-load orders, issued in June 2026 in place of a national rule; turbine shipment rates; and the next PJM auction.

THE NARROW PASS, AND ITS NEIGHBORS

TSMC, ASML, and the companies that wire the data center

Where it sits: the pass every barrel must go through, run by companies that don’t compete with their customers.

Why it matters: TSMC’s third-quarter revenue was $46.7 billion, up 41 percent, at a 66.4 percent gross margin. It announced another $100 billion for Arizona in July. It is reportedly planning price increases for 2027 and up to 20 percent more for CoWoS packaging, which TSMC won’t confirm. ASML says its planned supply for 2027 and 2028 “will not fully meet customer demand.” Around them, the less glamorous suppliers are growing faster than almost anyone. Celestica’s revenue rose 83 percent and it projects $40 billion for 2027. Amphenol rose 65 percent, Vertiv 61 percent and Lumentum 109 percent. Acquired’s hosts make the point about TSMC’s neutrality: it is “the only foundry at the leading edge that does not in any way compete with their customers.” Neutral suppliers get paid no matter which customer wins.

Watch: TSMC’s 2027 pricing, co-packaged optics (Coherent expects first revenue this quarter), and whether Celestica-style assemblers capture margin or get squeezed.

THE BANKERS

Private credit and the asset managers

Where it sits: Jay Cooke and J. P. Morgan, financing the railroads and reorganizing them when they fail.

Why it matters: a large and growing share of AI hardware is owned by financiers and leased to its users. Apollo led a $3.5 billion financing for GPUs leased to xAI in January and is arranging about $40 billion of GPU lease financing for SpaceX. Apollo and Blackstone built a $35 billion chip-lease vehicle for Anthropic, with Broadcom guaranteeing the chips’ residual value. Blackstone’s data-center platform was worth $185 billion in mid-2026. Blue Owl owns 80 percent of Meta’s Hyperion campus. Morgan Stanley has estimated that private credit could supply around $800 billion of data-center financing through 2028, a mid-2025 forecast. The twist is that AI also threatens these lenders: Blue Owl’s technology-lending fund faced redemption requests of about 39 percent of shares in the third quarter, capped at 5 percent, because investors fear AI will hurt the software companies it lends to.

Watch: redemption gates, GPU lease terms, and who ends up owning the hardware if a neocloud fails.

THE SOVEREIGNS

The Gulf, SoftBank, and national AI programs

Where it sits: Anglo-Persian in 1914 and OPEC in 1960: states buying in because the fuel is strategic.

Why it matters: the ambitions are enormous and the reality is smaller and riskier. Washington approved up to 35,000 GB300-class chips each for Saudi Arabia’s Humain and the UAE’s G42 in November 2025, roughly 50 to 70 megawatts apiece by rough estimate, against gigawatt-scale plans. Humain’s first Nvidia cluster starts in the fourth quarter of 2026, behind its original schedule. And the Gulf build-out is now physically exposed. In March 2026, drone strikes hit AWS facilities in the UAE and Bahrain, and Iran named G42, Nvidia, Oracle and Cisco as targets. Abu Dhabi’s MGX closed a $49 billion fund and invested in both OpenAI and Anthropic. SoftBank has committed about $64.6 billion to OpenAI. It funded that partly by selling its entire Nvidia stake and taking a $40 billion bridge loan due in March 2027, and S&P rates it BB+ with a negative outlook. Japan, France (Mistral raised €3 billion led by Samsung) and others are building national capacity on a smaller scale.

Watch: whether Stargate UAE’s first phase goes live, SoftBank’s refinancing, and new U.S. export approvals.

THE OTHER FIELD

China’s labs, chips and power

Where it sits: Baku and the Nobels, a parallel industry with its own supply, its own logistics and its own state backing.

Why it matters: Beijing, not just Washington, is now drawing the line. After the U.S. eased rules on Nvidia’s H200 in January 2026, Chinese customs reportedly told agents the chips were not permitted, and H200 sales to China came to under 1 percent of Nvidia’s data-center revenue last quarter. Huawei is promising its Ascend 960 in early 2027 and claims more than 1,000 large “SuperPoD” systems deployed. DeepSeek’s V4-Flash lists at $0.14 per million input tokens and says its Pro model’s price will fall once Huawei’s supernodes ship in volume. Zhipu and MiniMax listed in Hong Kong in January 2026; Moonshot has sought a $50 billion valuation. And China added roughly 540 gigawatts of power capacity in 2025, about eight times what the U.S. was projected to add. The binding constraint for Chinese AI is chips. For American AI, it is increasingly power.

Watch: Ascend production volumes, Chinese open-weight share on developer platforms, and the rare-earth decision on November 10.

THE GAS STATIONS

Apps, distribution and agent payments

Where it sits: the branded service station and the tank wagon, the last mile to the customer.

Why it matters: Apple has 2.5 billion active devices and spent only about $6.8 billion on capital expenditure in the first nine months of its fiscal year, less than each of Amazon, Alphabet and Meta spent in a single quarter. It rents Google’s Gemini for its next-generation models. That is a distributor collecting the margin without building the refinery. In enterprise software, Salesforce’s Agentforce passed $1.5 billion of annual recurring revenue, ServiceNow raised its AI target, and Palantir’s revenue grew 93 percent in the second quarter. And every agent purchase needs a payment rail. Stripe and OpenAI’s Agentic Commerce Protocol, Google’s AP2 and Universal Commerce Protocol, Visa’s and Mastercard’s agent programs, and Stripe’s Tempo network are competing to be the standard. Visa’s CEO calls agentic commerce “nascent,” and neither card network has disclosed agentic volumes.

Watch: the first disclosed agentic payment volumes, and whether Apple begins charging model makers for access to its users.

THE LANDHOLDERS

Data owners

Where it sits: the Pennsylvania farmers collecting royalties, still waiting for an OPEC.

Why it matters: the data owners are being squeezed from both sides. Reddit’s third-quarter revenue grew 53 percent, but it warned that search referral traffic “declined meaningfully” and its U.S. daily users fell. The WSJ reported it has weighed cutting off Google’s AI-training access as its roughly $60 million-a-year deal comes up for renewal. Cloudflare, which carries about a fifth of web traffic, blocks AI crawlers by default and offers pay-per-crawl. The courts have not settled the core question: in the UK, Getty lost its main copyright claim against Stability AI, and in the U.S., the New York Times case against OpenAI has no fair-use ruling or trial date. The content owners collecting real money are the ones with leverage, like Disney, which invested $1 billion in OpenAI while licensing its characters.

Watch: Reddit’s renewal terms, fair-use rulings, and whether gatekeepers like Cloudflare become the OPEC of content.

THE PARTS SUPPLIERS

Robot components

Where it sits: the suppliers to Ford: tires, glass, steel and spark plugs.

Why it matters: if robots are the new cars, their parts makers are the new auto suppliers. Hesai’s robotics lidar shipments rose 193 percent in the second quarter, and RoboSense’s robotics lidar rose 510 percent in the first half. Japan’s Harmonic Drive Systems and China’s Leaderdrive, which makes harmonic reducers for robot joints, report rising humanoid orders, small next to their totals for now. Underneath sits the most concentrated input of all: rare-earth magnets, where China mines about 69 percent of supply and processes up to 90 percent.

Watch: China’s rare-earth decision on November 10, 2026, and humanoid orders at the reducer makers.

THE RULE-MAKERS

Policymakers and regulators

Where it sits: Austin, Washington and Riyadh in the oil story: the people who decide who may produce, ship and sell.

Why it matters: nearly every chokepoint in this essay has a regulator attached. The Commerce Department’s export rules moved H200-class chips to case-by-case review in January 2026. FERC and the Energy Department are rewriting how large loads connect to the grid. Texas is implementing SB 6 by year-end and auditing its queue. The EU delayed its AI Act’s high-risk obligations to December 2027 but kept the rules for general-purpose models in force. China’s Commerce Ministry controls the rare-earth tap. The FCC holds SpaceX’s million-satellite application. In oil, the regulators arrived after the concentration. In AI, they are already sitting at every narrow point.

Watch: the rare-earth deadline, the FCC’s orbit ruling, and the next round of export rules.

The blind spots, in one table

BLIND SPOTSFacts from filings and reporting as of Oct 2026; “why underrated” is analysis
PlayerRoleWhy underratedSignal to watch
SK hynix, Samsung, MicronMemory (HBM, DRAM)Earning margins above Nvidia’s; SK hynix says it is sold out through 2029Contract prices; whether long-term-agreement price ceilings hold; HBM share shifts
GE Vernova, Siemens Energy, Mitsubishi HeavyGas turbines~226 GW contracted or reserved against roughly 25–30 GW shipped a year (rough estimate)Backlog vs. shipments; turbine prices; lead times
Transformer and switchgear makersGrid equipmentLead times up to four years; demand for some transformer types up 100–270% since 2019New U.S. factories coming online (Hitachi, Siemens)
PJM, ERCOT, state utility commissions, FERCPermission to connectPJM’s capacity auction cleared at its cap and still came up short; ERCOT’s large-load queue is ~475 GWTexas audit results; FERC large-load orders; capacity prices
TSMC and ASMLThe narrow passNeutral suppliers to everyone; ASML says 2027–28 supply won’t fully meet demandTSMC price increases for 2027; CoWoS capacity; ASML capacity expansion
Celestica, Amphenol, Arista, Vertiv, Lumentum, CoherentWiring the data centerGrowth of 37–130% in their latest quarters, far less attention than NvidiaCo-packaged optics adoption; 2027 order books
Apollo, Blackstone, Blue Owl, BrookfieldFinancing the railsThey own the GPUs in many deals; Blue Owl’s software fund faced 39% redemption requests on AI fearsRedemption gates; GPU lease pricing; defaults at the edges
Gulf sovereigns (Humain, G42, MGX)State-backed capacityAmbitions in gigawatts, approvals so far in tens of megawatts, and data centers now in a war zoneStargate UAE’s first 200 MW; new U.S. export approvals; security incidents
SoftBankOpenAI’s financierAbout $64.6B committed to OpenAI, funded partly with a $40B bridge loan due March 2027Refinancing the bridge; OpenAI IPO timing
Huawei, SMIC, DeepSeek, Moonshot, Zhipu, MiniMaxChina’s parallel stackBeijing blocked Nvidia’s H200 even after Washington allowed it; China added ~540 GW of power capacity in 2025Ascend 950/960 volume; DeepSeek pricing on domestic chips
Visa, Mastercard, Stripe, PayPal, ShopifyAgent payments and checkoutEvery agent purchase needs a payment rail and a merchant; volumes are still “nascent”First disclosed agentic payment volumes; protocol adoption (ACP, AP2, UCP)
AppleDistribution without capex2.5B active devices; spends a fraction of the hyperscalers’ capex and rents GeminiSiri usage; whether Apple charges model makers for access
Salesforce, ServiceNow, PalantirEnterprise agentsAgentforce ARR above $1.5B; Palantir revenue +93%AI revenue as a share of total; seat compression
Reddit, publishers, CloudflareData ownersLosing search traffic while renegotiating licenses; Cloudflare blocks AI crawlers by defaultReddit’s Google renewal; pay-per-crawl uptake; copyright rulings
Harmonic Drive, Leaderdrive, Hesai, RoboSense, magnet makersRobot partsRobotics lidar shipments up 193–510%; rare-earth magnets controlled by ChinaChina’s Nov 10 rare-earth decision; humanoid orders at reducer makers

ANALYSISThe common thread: the less-watched players sit at the physical and legal chokepoints (memory, turbines, transformers, grid permission, magnets, export licenses, orbital spectrum), or they own the hardware through finance. The 2026 data say these are where the scarcity and margins have been moving. That is the oil pattern exactly: when everyone watches the refiners, look at who owns the barrels, the rails and the permits.

SYNTHESISWhat held up

Part One’s core claim survives contact with the objections: value flows to the scarcest step, and the step keeps moving. What the objections add is a sharper picture of how it moves. Volume buys rebates, though in a shortage it buys priority and cheap credit more than discounts. Breadth buys time. Integration lets you take the margin wherever it is easiest to take. And the scarce step in 2026 is not one thing. It is chips for some buyers, memory for others, power and turbines for everyone, and launch, if orbit ever becomes the cheapest place to compute.

The castings, updated

CASTINGS, PART ONE VS PART TWOAnalysis
PlayerPart One castingPart Two castingWhy it moved
Nvidia / Jensen HuangVanderbilt: the toll on everything that movesVanderbilt who also lends shippers the money for freight cars$279B of supply commitments, capacity backstops, up to $105B of lease guarantees, stakes in its customers
SpaceX / xAI / MuskRail baron crossed with the NobelsVanderbilt with a refinery, building the only line to orbitThe Super Intelligence segment, mostly compute rented to Anthropic, Google and Reflection AI, was 52% of Q3 2026 revenue; 80%+ of the world’s mass to orbit
GoogleStandard Oil (structurally)Standard around 1901, not 1911: dominant, facing new supply, decline optionalSearch growth slowed to 12%, but revenue +25%, Cloud +103%, TPUs, 1B Gemini users
MetaBarely castThe integrated major that burns its own refined productOwn data centers, closed Muse models, AI-tuned ads at $63B a quarter; 97.6% ads
OpenAIStandard’s brand without the pipelinesUnchanged; the ad business is real but small~$1B ad run rate vs Google Search at ~$253B annualized; enterprise now the majority
AnthropicA big refiner paying the railroadFlagler’s strategy without Flagler’s ownershipPlays Google, Amazon, Microsoft, SpaceX and neoclouds off each other; revenue $16.8B in Q3
OracleA railroadJay Cooke’s Northern Pacific more than the Pennsylvania’s Empire$664B backlog, about half OpenAI; BBB-; stock down ~62% from peak
NeocloudsRailroads at the edgeCleveland’s refiners before 1872: numerous, leveraged, ripe for consolidationCoreWeave CDS over 1,100bp; Nscale’s top customer 52% of revenue
Memory and turbine makersNot castThe 1861 barrel: briefly worth more than the oilMicron 86.8% gross margin; ~226 GW of gas turbines contracted or reserved at the big three
Waymo / Tesla / UberThe car, in generalFirst refinery with a product / Ford’s factory / the service station500k+ paid rides a week; cheapest cost per mile (est.); aggregator of everyone’s fleets
AmazonA railroad (AWS)Also Standard’s tank wagons: the fenced last mile of shoppingAlexa for Shopping at 400M users; 99 outside agents blocked

The hypotheses, judged

HYPOTHESIS SCORECARDAnalysis — verdicts as of Oct 31, 2026
HypothesisVerdictWhy
Orbital data centers deserve to be taken seriouslyHOLDSSpaceX’s own filing, a named Nvidia-based satellite, rivals at Google, Blue Origin and in China. Part One undersold it.
Launch becomes the new railroad, with an Nvidia-like moatPARTLYThe launch lead is durable, plausibly more than Nvidia’s. Orbital compute itself only wins if launch gets ~10x cheaper; until then it’s an option on a niche.
Robots are AI’s gasoline momentPLAUSIBLEThe market they attack (car ownership, a ~$2.6T auto industry) is the right size, and unit costs are falling fast.
…and underweighted outside a small groupHALFPublic attention lags, but capital doesn’t: Waymo at $126B, Unitree’s 629% debut, a Tesla pay plan built on robots.
…with Tesla/SpaceX best placed and Waymo less commercializedDOESN’T HOLD TODAYWaymo is far ahead commercially; Chinese makers lead humanoid volume and vehicle cost. Tesla’s edge is manufacturing cost, a long-run bet.
Meta is underweightedHOLDSOwns infrastructure, has closed frontier models and the strongest AI-for-ads results; caveat: ads are 97.6% of revenue.
The interface future is more Neuralink than glassesNOT THIS DECADEDozens of implants vs millions of glasses; Meta’s EMG wristband is the middle path, and it’s shipping.
Meta’s hardware record is weaker than Musk’sHOLDSPortal, Quest and ~$96B of Reality Labs losses vs Tesla’s and SpaceX’s manufacturing scale; glasses are the exception, made by a partner.
Google is Standard after its primePARTLYSearch decelerating and AI Mode monetizing less, but a 25%-growth company with chips, cloud and distribution. 1901, not 1911.
OpenAI is best placed to dominate AI adsDOESN’T HOLD (YET)Its ad business is ~0.4% of Google Search; Google and Meta already sell ads inside AI; OpenAI’s commerce push retreated.
Hyperscalers are railroads; volume-controlling labs are StandardPARTLYRight about contracts and rebates (price, priority, credit); wrong so far about margins, which sit with Nvidia and the clouds.
Someone consolidates the neocloudsLIKELY, NOT YETNo verified 2026 merger; the first round is more likely through lenders, anchor tenants and Nvidia than a single buyer.
Oracle is the Pennsylvania Railroad overplaying its handPARTLYOverextension fits; the Empire move (infrastructure into the customer’s business) is being made by stronger players.
Elon is RockefellerAMBITION, NOT POSITIONIntegration is real and growing; low-cost dominance in models is not.
Revenue breadth is survivalHOLDSThe fragile layer is single-revenue, debt-financed infrastructure, exactly where the stress shows.
Pure token production loses value fastHALFCommodity tokens yes; frontier tokens hold price and grow in volume.
AI runs oil-style boom-bust cyclesHOLDSGPU rents fell ~80% then rose 44%; DRAM nearly doubled in a quarter; the scarce input moved from GPUs to memory and power.
Consumer AI settles at $50–100 a monthAS TOTAL MONETIZATION, NOT PRICEFew pay even $20 today; ads plus subscriptions plus commerce could reach it per engaged user.
Amazon can win agentic shoppingLEANS YESIts own agent already has 400M users and ads; it fences outsiders out, though courts have weakened the legal fence.
Nvidia can move across the stackHOLDS, AS FINANCIERSystems, networking, open models and above all finance; deliberately not a rival cloud or lab. New risk: its own guarantees.

Three patterns run through the verdicts. First, the hypotheses about direction mostly held: orbit matters, robots are the likeliest gasoline moment, Meta was underweighted, the chain shifts, breadth protects. Second, the hypotheses about who is ahead right now mostly didn’t. Waymo, not Tesla, leads robotaxis. Google, not OpenAI, still sells the ads. Nvidia and the clouds, not the volume-controlling labs, keep the margin. Third, the riskiest place in the chain is the same place Part One flagged: debt-financed infrastructure with one big customer. It now includes the balance sheets of the suppliers, Nvidia’s above all, that have guaranteed it.

What to watch

  1. Starship refueling and the first Starmind launch. SpaceX plans an on-orbit refueling attempt in early 2027 and Starmind flights in 2027. If both happen on time, scenario 1 gets real. If they slip, orbit stays a niche.
  2. The FCC’s ruling on the million-satellite application. It is the first test of whether a Railroad Commission comes to orbit.
  3. China’s rare-earth decision on November 10, 2026. The suspension of its broad export controls expires that day. An extension keeps robots and EVs supplied; a lapse puts a chokepoint under the gasoline moment.
  4. Waymo’s next ride count and Tesla’s first disclosed fleet numbers. A million paid rides a week was Waymo’s year-end goal. Tesla has never disclosed paid rides.
  5. Google’s fourth-quarter search growth and AI Mode ad pricing. Another step down from 12 percent would strengthen the “after its prime” case.
  6. OpenAI’s ad revenue against its $2.5 billion 2026 target. The gap between a run-rate claim and booked revenue will be visible in its eventual filings.
  7. Neocloud credit. CoreWeave’s CDS, Oracle’s rating, Lambda’s and Nscale’s IPOs. A failed raise is the Cleveland moment.
  8. Memory and GPU rental prices. If H100 rents keep rising while DRAM stays tight, the cobweb is still on its upswing. A sharp reversal is the next glut.
  9. Nvidia’s guarantees. The size of its supply commitments, backstops and lease guarantees relative to revenue is now as important as its gross margin.
  10. Agent access law. The Amazon–Perplexity case, back in district court, decides whether stores can fence out shoppers’ agents.

The short version

Rockefeller won by controlling volume, bargaining it into rebates, and then integrating until there was nothing left to bargain with. In AI, everyone is now trying to do some version of that at once. The labs promise volume. The landlords promise capacity. Nvidia and the memory makers demand prepayments and then lend the money back. SpaceX rents its rails to rivals while it builds a line to orbit. Google and Meta try to own every layer.

The evidence of late 2026 says the integrators and the toll collector are winning, the debt-financed middle is the most fragile, and the gasoline moment is most likely to arrive on wheels and legs rather than in a chat box. Orbit is the wild card: a real railroad if launch gets cheap enough, a niche if it doesn’t.

Own a scarce step, sell to everyone, and don’t borrow against a single customer’s promises.

START FROM THE BEGINNINGRead Part One: If AI Is the New Oil, Who Is Rockefeller? →

This essay is analysis and opinion built on dated, sourced figures as of October 31, 2026. Company run-rate claims, forecasts and third-party estimates are labeled as such; where sources disagree, the text says so. It is not investment advice; mentions of companies are not recommendations.

DON'T TAKE MY WORD FOR IT

Research trail

Dwarkesh Podcast — Elon Musk (with John Collison), Feb 5 2026 ↗Dwarkesh Patel — Notes on Space GPUs ↗WEF Davos 2026 — Elon Musk with Larry Fink (transcript) ↗Acquired — Meta ↗Acquired — The Mark Zuckerberg Interview ↗Acquired — Google: The AI Company ↗Acquired — Alphabet Inc. ↗Acquired — Nvidia Part II: The Machine Learning Company ↗Acquired — Nvidia Part III: The Dawn of the AI Era ↗Acquired — Jensen Huang interview ↗Acquired — TSMC founder Morris Chang ↗Acquired — Standard Oil Part I ↗Acquired — Standard Oil Part II ↗SpaceX — Q3 2026 earnings release (SEC) ↗SpaceX — S-1/A, June 2026 (SEC) ↗SpaceX — Q3 2026 earnings call transcript ↗SpaceX — Q2 2026 earnings call transcript ↗NBC News — SpaceX revenue jumps on data-center rentals ↗FCC — Public Notice DA 26-113 (SpaceX orbital data center system) ↗CNBC — FCC chair on Amazon’s petition against SpaceX ↗Via Satellite — Amazon petition to deny SpaceX constellation ↗WilmerHale — The legal reality of orbital data centers ↗Via Satellite — FCC adopts five-year deorbit rule ↗Google Research — Project Suncatcher ↗GPB/AP — Google launches Suncatcher prototype ↗GPB/AP — Starship’s first orbital mission ↗TechCrunch — Blue Origin enters the space data center game ↗SiliconANGLE — Starcloud raises $250M ↗CGTN — China launches satellites for space-based computing ↗IEEE Spectrum — The economics of orbital data centers ↗Forethought — Will we really put data centers in space? ↗Power Engineering — Gas turbine prices climb 195% ↗Rest of World — Orbital data centers and AI sovereignty ↗Futurism — Astronomers on SpaceX’s million-satellite plan ↗InsideEVs — Waymo doubles weekly rides ↗electrive — Waymo raises $16B ↗Waymo blog ↗The Charge Port — robotaxi tracker ↗Tesla Q3 2026 earnings call transcript ↗Drive Tesla Canada — Tesla Q3 2026 results ↗Not a Tesla App — Optimus production begins ↗CnEVPost — China suspends new robotaxi permits ↗Baidu — Q2 2026 results (SEC) ↗Pony.ai — Q2 2026 results (SEC) ↗WeRide — Q2 2026 revenue ↗SSE — Unitree IPO disclosure ↗Caixin — Unitree soars in Shanghai debut ↗TechNode — China’s share of humanoid shipments, H1 2026 ↗Boston Dynamics — new Atlas ↗Motor1 — Zoox exemption and paid rides ↗Sherwood — Robotaxi costs per mile (Morgan Stanley) ↗TechCrunch — Waymo vs Uber price gap ↗AAA — Cost of new-vehicle ownership 2025 ↗EIA — U.S. gasoline consumption 2025 ↗Yahoo Finance — China rare-earth export pause nears expiry ↗Meta — Q3 2026 results (SEC) ↗Facebook — Q4 2013 results (SEC) ↗Crypto Briefing — Meta’s $95B equity raise ↗TechCrunch — Meta releases Muse Wonder ↗Fortune — Meta unveils Muse Spark ↗Global Data Center Hub — Hyperion scales to 5 GW ↗IDC — Smart glasses growth cools ↗ChannelNews — EssilorLuxottica Q3 2026 ↗The Register — Vision Pro and Quest sales drop ↗TechCrunch — Neuralink’s $650M Series E ↗Paradromics — first human implant ↗MedTech Dive — Precision Neuroscience FDA clearance ↗Alphabet — Q3 2026 results (SEC) ↗Alphabet — Q1 2026 results (SEC) ↗Alphabet Q3 2026 earnings call transcript ↗Google — Sundar Pichai at I/O 2026 ↗StatCounter — search engine market share ↗PPC Land — Google’s ad-tech remedies ↗PPC Land — Google appeals search remedies ↗NBC News — OpenAI starts testing ads in ChatGPT ↗MarketScreener/Axios — OpenAI ad revenue projections ↗TechAfrica News — ChatGPT ads at $1B run rate ↗AI Weekly/eMarketer — chatbot ad revenue forecast ↗Retail Systems — OpenAI pivots from Instant Checkout ↗The Decoder — OpenAI shuts Atlas browser ↗Ahrefs — ChatGPT vs Google search volume ↗TechCrunch — ChatGPT at 900M weekly users, 50M subscribers ↗The Verge — ChatGPT at 1.2 billion weekly users ↗TechTimes — OpenAI enterprise revenue tops consumer ↗CNBC — OpenAI revenue and the Oracle/CoreWeave selloff ↗CoreWeave — Q2 2026 results ↗CoreWeave — $8.5B investment-grade GPU financing ↗CoreWeave — 8-K on Nvidia capacity agreement ↗CoreWeave — Core Scientific stockholder vote ↗Investing.com — AI credit default swap spreads widen ↗Nebius — Q2 2026 shareholder letter (SEC) ↗TechCrunch — Lambda to raise $4B ↗Zawya — Crusoe valued at $30.9B ↗Webull/Barchart — Nscale S-1 ↗Forbes — Fluidstack worth $18B ↗TeraWulf — Google backstop of Fluidstack leases (SEC) ↗ERP Today — Oracle Q1 FY27 results ↗PPC Land — S&P cuts Oracle to BBB- ↗Stock Analysis — Oracle ↗Nvidia — 10-Q, quarter ended July 26 2026 (SEC) ↗Nvidia — financing platforms with Apollo, BlackRock and others ↗Bloomberg Law — Huang rules out $100B OpenAI investment ↗SDxCentral — Nvidia pulls back on DGX Cloud (report) ↗Nvidia — Nemotron models ↗Nvidia — Nebius partnership ↗W.Media — Nvidia licenses Groq technology ↗Broadcom — Q3 FY26 results (SEC) ↗Micron — FQ4 2026 results (SEC) ↗Micron — FY2026 10-K (SEC) ↗TrendForce — Samsung, SK hynix long-term agreements ↗TrendForce — DRAM contract prices, 1Q26 ↗Sovereign Magazine — How long will the memory shortage last? ↗IntuitionLabs — H100 rental prices compared ↗Ornn — H100 compute price index ↗Silicon Data — H100 price spike ↗FourWeekMBA — H100 rent at $2.88 (Silicon Data) ↗ColoPrice — H100 rental price history ↗Gigazine — Anthropic and SpaceX’s Colossus deal ↗Anthropic — Higher limits with SpaceX capacity ↗CNBC — Anthropic third-quarter revenue ↗Mostly Metrics — Anthropic S-1 breakdown ↗SEC — Anthropic PBC Form S-1 ↗Meta — Hyperion joint venture with Blue Owl ↗Bloomberg Law — Blue Owl halts redemptions ↗Amazon — Q3 2026 results (SEC) ↗Amazon — Q4 2025 results ↗Microsoft — FY27 Q1 results ↗Apple — fiscal Q3 2026 results (SEC) ↗Storage Newsletter — Nvidia Q2 FY27 results ↗Storage Newsletter — Oracle Q1 FY27 results ↗TechCrunch — ChatGPT’s $100 Pro plan ↗Claude — pricing ↗Google One — AI plans ↗Menlo Ventures — 2025 State of Consumer AI ↗Deloitte — Digital Media Trends dashboard ↗Benton — Cost of staying connected 2026 ↗Epoch AI — LLM inference price trends ↗Epoch AI via P. Pillitteri — the plunging price of thought ↗AWS — Bedrock GPT-5.6 Sol pricing ↗Amazon — statement on Perplexity’s Comet ↗Engadget — Ninth Circuit vacates Amazon injunction ↗Amazon — robots.txt ↗Search Engine Journal — Amazon blocks Meta’s Muse ↗Bloomberg Government — Amazon and OpenAI ChatGPT ads partnership ↗Amazon — AWS and OpenAI $38B agreement ↗Digital Commerce 360 — McKinsey agentic commerce forecast ↗