COBURNICUSThe film starts with a game. I hand you ten billion dollars. It’s fake money, but pretend it’s real. Your job is to build the next great AI, and you have an idea nobody at the big labs has tried. All you need now is the machines.
Then you find out how far your fortune goes. Amazon, Alphabet, Microsoft and Meta guided to roughly $730 billion of capital spending for 2026, which works out to about two billion dollars a day. Your ten billion buys about five days of their budget.
The money turns out to be the smaller problem. The chips, the memory, the turbines and the place in line at the power company have mostly been promised to somebody else already, often for years. The biggest companies on Earth got to the front of every line at once, and they’re borrowing whatever it takes to stay there. I call that the compute wall.
I made this film because I wanted to know whether a newcomer can still get through it. So the film makes four phone calls.
Call one: the chips
Start with who owns what already. Epoch AI estimates the world had about twenty million H100s’ worth of AI computing power at the end of 2025, and that five companies (Amazon, Google, Meta, Microsoft and Oracle) held 71 percent of it, up from 63 percent in early 2024. Google alone held about a quarter.
Some of those chips aren’t even plugged in. On the BG2 podcast in October 2025, Microsoft CEO Satya Nadella was asked about a possible compute glut. His answer: “you may actually have a bunch of chips sitting in inventory that I can’t plug in. In fact, that is my problem today.” He said the shortage was warm shells, meaning buildings with power ready. Microsoft kept buying anyway.
A modern AI chip needs a processor, stacks of high-bandwidth memory (HBM) and an advanced packaging step called CoWoS that bonds them together. Each has its own queue. SK Hynix told investors in October 2025 that its 2026 supply was sold out. In July 2026 its CEO told Reuters that 2027 will be the worst year for memory supply in the industry’s history. Epoch found that the four biggest AI chip designers took around 90 percent of the world’s CoWoS and HBM supply in 2025.
Older chips should be the cheap way in. Instead, SemiAnalysis tracked one-year H100 rental prices rising from about $1.70 an hour in October 2025 to about $2.35 by March 2026, almost 40 percent in five months. I’ve spent my career watching hardware lose value faster than a new car, and I have never seen that happen with a three-year-old processor.
Call two: the power
Say a truck full of chips shows up. Now you need electricity. Lawrence Berkeley National Lab’s Queued Up 2026 report puts the median wait for power projects that came online in 2025 at 61 months from application to operation. In Texas, the queue of huge new customers asking ERCOT for power went from 63 to 226 gigawatts in about a year.
So you skip the grid and build your own gas plant. GE Vernova reported more than 120 gigawatts of gas turbines on order or reserved in October 2026, and its CEO said in July that it was taking reservations for 2031 delivery.
Even OpenAI gets rationed. In April 2026, Oracle said 200 megawatts were running at the Stargate campus in Abilene, a site designed for about 1.2 gigawatts. Epoch estimates it reached full build only at the start of November.
Calls three and four: the landlords and the bank
Renting used to be the way around all this. CoreWeave said in March 2026 that its 2026 capacity was broadly sold out, with take-or-pay contracts stretched from about three years to mostly five, some six. You pay whether you use the machines or not.
Borrowing has its own wall. The five biggest cloud spenders issued $121 billion in US bonds in 2025, against an average of $28 billion a year from 2020 to 2024, according to Bank of America figures. When Alphabet borrowed in February 2026, it paid roughly a quarter point to one point over Treasuries. A startup has no credit rating at all.
The film also walks through the accounting arguments. Michael Burry, who was betting against AI stocks at the time, called understated depreciation “one of the more common frauds of the modern era” in November 2025. That’s his accusation. I couldn’t find any regulator or court that has found these companies committed accounting fraud. Nvidia’s CFO says A100s it shipped six years ago still run at full utilization, and Amazon shortened some server lives from six years to five. Nobody knows yet who’s right.
Who gets through
New labs still get in, and how they get in is the story. Thinking Machines, founded by former OpenAI CTO Mira Murati, agreed with Nvidia in March 2026 to deploy at least a gigawatt of Nvidia’s next-generation chips. Jensen Huang puts a gigawatt at $50 to 60 billion. Reflection, an open-model startup, will pay SpaceX’s AI arm (the former xAI) $150 million a month through 2029 for chips at Colossus 2, Axios reported, and after the first three months either side can exit on 90 days’ notice.
That’s why I keep coming back to the word permission. A chipmaker, a cloud company or a billionaire now has to decide you’re worth a slot.
Two things that didn’t make the cut
The renting goes both ways at the top. Apollo noted in June 2026 that Anthropic and Google both signed deals to rent capacity from xAI. CNBC reported Google’s deal at $920 million a month from October 2026 to June 2029, and the reported Anthropic deal is $1.25 billion a month for Colossus 1 through May 2029. Google is the single largest owner of AI compute, by Epoch’s count, and it’s still renting more from a rival.
Household bills come into it too. PJM’s capacity price, a charge that flows into electric bills across the mid-Atlantic, went from $28.92 to $329.17 per megawatt-day, according to IEEFA, and PJM’s market monitor attributed 63 percent of the 2025/26 increase to data centers. That helps explain why towns from Charlotte to Fort Worth have voted on data center moratoriums.
The open question
Critics make a real case against the wall. DeepSeek released V4 with open weights in April 2026 under export controls, and its own report puts it three to six months behind the closed frontier. Analysts expect TSMC’s packaging capacity to double by 2028. If the boom goes bust, cheap compute could flood the market, the way cheap fiber did after the dot-com crash.
One test is whether a lab with no chipmaker or cloud giant behind it reaches the frontier. Will that happen before the current contracts, some of which run through 2032, decide who’s left standing? I hope the next great AI comes from somewhere nobody’s looking. When it does, look closely at the announcement, and check whose name is on the lease.
Sources
- Epoch AI, hyperscalers control most compute (2026): https://epoch.ai/data-insights/hyperscalers-control-most-compute
- Epoch AI, frontier labs don’t use most AI compute (May 2026): https://epoch.ai/gradient-updates/frontier-labs-dont-use-most-ai-compute
- Epoch AI, AI chip supply chain constraints (Mar 2026): https://epoch.ai/data-insights/ai-chip-supply-chain-constraints
- BG2 Pod with Sam Altman and Satya Nadella (Oct 2025): https://www.youtube.com/watch?v=Gnl833wXRz0
- Reuters on SK Hynix CEO’s 2027 supply warning (Jul 2026): https://money.usnews.com/investing/news/articles/2026-07-10/sk-hynix-ceo-sees-worst-ever-memory-supply-shortage-in-2027-says-demand-to-outstrip-supply-beyond-2030
- SemiAnalysis, The Great GPU Shortage (Apr 2026): https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity
- LBNL, Queued Up 2026: https://eta-publications.lbl.gov/sites/default/files/2026-06/queued_up_2026_edition.pdf
- GE Vernova Q3 2026 results: https://www.sec.gov/Archives/edgar/data/0001996810/000199681026000162/gev3q2026form8-k10282026.pdf
- Epoch AI, Stargate Abilene status: https://epoch.ai/data/ai-data-centers/directory/openai-stargate-abilene
- CoreWeave contract terms (Mar 2026): https://finance.yahoo.com/news/coreweave-conference-insatiable-ai-demand-040315348.html
- Reuters on hyperscaler bond sales (Jul 2026): https://kfgo.com/2026/07/29/hyperscaler-debt-binge-pushes-yields-up-as-investor-demand-cools/
- Michael Burry on depreciation (X, Nov 2025): https://x.com/michaeljburry/status/1987918650104283372
- CNBC on Nvidia and Thinking Machines (Mar 2026): https://www.cnbc.com/2026/03/10/nvidia-mira-murati-thinking-machines-lab-ai.html
- Axios on Reflection and SpaceX (Jun 2026): https://www.axios.com/2026/06/22/open-source-ai-gets-more-compute-from-spacex
- Apollo, The Growing Compute Shortage (Jun 2026): https://www.apollo.com/wealth/insights-news/insights/2026/06/growing-compute-shortage
- CNBC on Google renting xAI compute (Jun 2026): https://www.cnbc.com/2026/06/05/google-to-pay-spacex-920-million-a-month-for-xai-compute-capacity.html
- IEEFA on PJM capacity prices: https://ieefa.org/resources/projected-data-center-growth-spurs-pjm-capacity-prices-factor-10
- DeepSeek V4 technical report: https://arxiv.org/html/2606.19348v1
Subscribe to get the next film and the research behind it. All the films are on YouTube at CoburnicusMind.