Last Updated on July 6, 2026 by Taya Ziv
Last updated: July 7, 2026. This is a living pillar. We update it as the war moves, and we log every change at the bottom of the page.
Quick answer: The AI infrastructure wars are the 2026 shift in AI competition away from who has the best model and toward who controls the supply chain underneath the models: chips, compute capacity, electricity, and capital. Roughly a dozen companies and at least two governments now decide who gets to run AI at scale and at what price, and if you build anything on top of AI, they set your costs, your margins, and your ceiling.
Between May 6 and June 5 of this year, I published nine stories that looked, at the time, like nine separate news items. An AI lab renting an entire data center from a rocket company. A chipmaker investing tens of billions into its own customers. A startup dropping its servers into the ocean. A battery manufacturer writing a $740 million check into one of the largest AI funding rounds in history. A software giant firing its most famous partner on a conference stage, to applause.
I covered them one at a time, like a reporter. That was the mistake, and I’ll own it: I write about this industry for a living and it still took me weeks to see it. Pin those nine headlines to a wall, run a string between them, and you’re not looking at nine stories. You’re looking at one story wearing nine disguises.
The story is this: the AI industry stopped competing on models and started competing on everything underneath them. And almost every founder I talk to is still watching the old war.
The war nobody bothered to declare
For two years, the AI conversation was a model conversation. GPT versus Claude versus Gemini versus DeepSeek. Benchmarks, context windows, reasoning scores. Every launch day was a title fight, and we all scored the rounds.
That fight isn’t over, exactly. But it stopped being the fight that decides anything. Because somewhere around the start of 2026, the frontier labs quietly discovered the same thing at the same time: the best model in the world is worthless if you can’t run it. Compute became the constraint. Power became the constraint on compute. Capital became the weapon for securing both. And control of that stack, not model quality, became the thing the giants actually fight about.
The clearest single piece of evidence came on May 6. Anthropic, fresh off usage growth of 80x in a single quarter, needed more compute than every deal it had already signed could provide. So it went out and rented the biggest GPU cluster on the market, all 220,000 GPUs in Elon Musk’s Colossus 1 data center in Memphis, more than 300 megawatts of capacity, in one deal with SpaceX. Two weeks earlier, this same company had walked away from classified Pentagon work on principle. Then it signed a landlord agreement with the most polarizing man in tech, who cheerfully reported that “no one set off my evil detector.”
Tom Brown, Anthropic’s co-founder and chief compute officer, explained the deal in one sentence that should be taped above every AI founder’s desk: “We’re going to need to move a lot of atoms to keep up with AI demand, and there’s nobody better at quickly moving atoms, on or off planet Earth.”
Atoms. Not tokens, not parameters. Atoms. The most advanced software companies on the planet are now bottlenecked by steel, silicon, and electricity, and they are behaving accordingly. That’s what an infrastructure war looks like from the inside: ideology bends, rivals become landlords, and the press release calls it a partnership.
Front one: compute became the moat, and the moat is rented
Here’s the uncomfortable arithmetic of 2026. Model quality converged. Training know-how leaked into the open literature. Talent moves between labs every eighteen months. The one thing that did not converge, cannot leak, and does not change employers is access to physical compute.
Which means the competition reorganized itself around that scarcity. Whoever can secure GPUs at scale wins the right to serve demand. Whoever can’t, queues. And the queue is not neutral, because the company selling the chips is also picking who gets them.
That’s not an insinuation. It’s documented behavior. This year, Nvidia crossed $40 billion in AI equity investments in five months, including a $30 billion stake in OpenAI, its own largest customer, plus $3.2 billion into Corning, which makes the optical guts of its data center cables, and $2.1 billion into IREN, a data center operator that runs Nvidia hardware. That’s on top of 67 venture deals in 2025. The pattern is consistent: nearly every company Nvidia funds is also a company that buys from Nvidia, and the checks often come paired with capacity reservations and roadmap agreements. Wedbush analyst Matthew Bryson acknowledged the circularity and argued it could still build Nvidia a durable competitive moat. He’s probably right, which is exactly the problem. When the arms dealer picks sides, the sides were never equal.
So compute is the moat now. But notice something strange about this moat: almost nobody owns it outright. Anthropic rents from SpaceX. OpenAI’s capacity is a lattice of commitments across Microsoft, Oracle, and whoever else can pour concrete fast enough. The moat is leased, and the leases have landlords, and the landlords have opinions.
Front two: the money went circular
Follow the capital and the war gets clearer, and weirder.
Start with the venture side. Peter Thiel’s Founders Fund deployed $4.6 billion into just seven companies in eleven months, then raised a fresh $6 billion, its largest fund ever, to do it again. Look at the names: Anthropic, Anduril, SpaceX, OpenAI, Stripe, Ramp, Cognition. Every single one is a platform or infrastructure play. Not one application company. Not one bet on the layer where most founders actually live. And this is not one eccentric fund; it’s the whole market’s center of gravity, the same concentration we saw when roughly $300 billion of venture capital landed in Q1 2026 and four companies absorbed a terrifying share of it.
Now look at what happens when an infrastructure challenger tries to stand up on its own. Cerebras, the wafer-scale chip company everyone crowned the “Nvidia killer,” went public at a $48.8 billion valuation with an order book 20 times oversubscribed. Sounds like independence. Except that buried in the S-1, OpenAI holds warrants for 10 percent of Cerebras at an exercise price of $0.00001 per share, essentially free, attached to the $20 billion deal that justifies the valuation. Two customers from a single country generated 86 percent of its 2025 revenue. The challenger isn’t a challenger. It’s a captive supplier with a ticker symbol.
So run the loop. Nvidia invests in OpenAI. OpenAI spends the money on Nvidia chips and takes 10 percent of Nvidia’s competitor as the price of an order. Nvidia’s revenue rises, its stock rises, and the larger market cap funds the next round of investments into its own customers. Money goes out the door, walks around the block, and comes home wearing a different hat. Wall Street calls this an ecosystem. My read: it’s a market grading its own homework.
I want to be fair here: circular financing is not automatically fake demand. The GPUs are real, the usage growth is real, Anthropic’s 80x quarter is real. But when the buyer, the seller, and the financier keep turning out to be the same four entities, price signals stop meaning what founders think they mean. You cannot read this market the way you’d read a normal one.
Front three: the grid said no
While the money spins in circles, physics has been quietly vetoing everyone’s plans.
The International Energy Agency projects global data centers will consume 1,000 terawatt-hours of electricity in 2026, roughly Japan’s entire national consumption. It still isn’t enough. Nearly half of the data center projects scheduled to finish this year are delayed, not for chips or money, but because the grid cannot deliver the watts. Morgan Stanley forecasts a 49 gigawatt power shortfall in the US alone by 2028, and lead times for high-voltage transformers stretched from 18 months to almost four years. Bloomberg confirmed that OpenAI’s Stargate campus, the largest AI infrastructure announcement in history, was still an empty field this spring. The biggest AI company on earth announced a megaproject, and the electrical grid shrugged.
This is why the most telling deal of the whole cycle might be the smallest one on the board. A Portland startup called Panthalassa raised $140 million, led by Thiel, to float wave-powered data centers in the open ocean, generating power at a claimed $0.02 per kilowatt-hour with free seawater cooling. No grid connection, no land permits, no transformer queue. “Extra-terrestrial solutions are no longer science fiction,” Thiel said of the raise. “Panthalassa has opened the ocean frontier.”
Read that again and appreciate how absurd 2026 actually is. The constraint on artificial intelligence, the most software of all software stories, is now so physical that serious money is going into steel tubes bobbing in the sea. Meanwhile Anthropic’s SpaceX agreement includes exploratory talks about compute in orbit. When your industry’s growth plan involves the ocean and outer space because the land is full, the bottleneck is not a rumor.
Front four: flags on the map
Then there’s the front that worries me most, because nobody involved is even pretending it’s business.
DeepSeek spent two years as the industry’s favorite counterexample: a Hangzhou lab funded by hedge fund profits, no outside investors, shipping open-weight models that undercut American frontier pricing by roughly 90 percent. In May, that story ended. China’s state-run Big Fund, the vehicle Beijing uses to steer its semiconductor industry, moved to lead DeepSeek’s first outside round at a valuation north of $45 billion. By June the round had swollen into something stranger: $7.4 billion at a valuation between $52 and $59 billion, led by the founder’s own $2.8 billion check alongside Tencent, battery giant CATL, NetEase, JD.com, and China’s national AI fund, with not one traditional venture firm on the list.
When the investors in an AI round make batteries, video games, and groceries, you are not reading a term sheet. You are reading industrial policy. CATL isn’t chasing model benchmarks; it makes the storage systems that keep data centers alive on renewable power, and it’s locking in its relationship with the demand source. The state fund isn’t seeking returns; it’s planting a flag on a company that embarrassed Silicon Valley. Sovereign capital has entered the compute war, and sovereign capital does not exit, does not mark to market, and does not lose interest next cycle.
The American version is less centralized but not that different in effect: a chip vendor, three hyperscalers, and a handful of mega-funds allocating capital by thesis rather than by market price. Two systems, one behavior. Compute is being treated the way nations treat oil, and pricing follows politics.
Front five: even the allies are arming
And in case anyone still believed the era’s partnerships were load-bearing, Microsoft spent this spring demonstrating what a $100 billion friendship is worth.
First came the quiet part. In sworn testimony during the Musk trial, a Microsoft executive put the company’s total OpenAI spend above $100 billion, against roughly $9.5 billion in recognized revenue coming back. Around the same time, Satya Nadella’s internal framing leaked into the reporting, one sentence that explains everything Microsoft has done since: “I don’t want to be IBM and OpenAI to be Microsoft.” Hence the shopping trip we covered, when Microsoft started quietly evaluating replacement models, including the diffusion-based startup Inception.
Then came the loud part. At Build 2026, Microsoft unveiled Project Polaris and announced it will replace GPT-4 Turbo as GitHub Copilot’s default engine starting August 2026, running on Microsoft’s own Maia accelerators, alongside an entire in-house model family covering reasoning, image, voice, and transcription. Own silicon, own models, own cloud, own distribution. The deepest alliance in AI ended as a market research program with a very expensive tuition bill.
If Microsoft, holding the strongest hand at the table, refused to depend on a single AI supplier, what exactly is your plan for depending on one?
The ledger: sixty days that redrew the map
Here’s the full board in one table. Every number below comes from our own reporting in the linked spokes, which carry the original sourcing (IEA, Morgan Stanley, Bloomberg, S-1 filings, trial testimony).
| Covered (2026) | Who moved | The move | The number that matters | |—|—|—|—| | May 6 | Beijing’s Big Fund | Talks to lead DeepSeek’s first outside round | $45B+ valuation | | May 6 | Anthropic × SpaceX | Rented all of Colossus 1 in Memphis | 220,000 GPUs, 300+ MW, after 80x usage growth | | May 7 | Founders Fund | Closed its largest fund ever, all infrastructure bets | $6B new, after $4.6B into 7 companies in 11 months | | May 11 | Nvidia | Equity into its own customers | $40B in 5 months, $30B to OpenAI alone | | May 12 | The power grid | Half of 2026 data center projects delayed on power | 1,000 TWh demand (IEA), 49 GW US shortfall by 2028 (Morgan Stanley) | | May 12 | Panthalassa | Wave-powered ocean data centers | $140M raise, $0.02/kWh claimed cost | | May 13 | Cerebras / OpenAI | “Independent” chip challenger IPO | $48.8B valuation; OpenAI warrants for 10% at $0.00001/share | | May 15 | Microsoft | Revealed OpenAI spend vs. return, began shopping | $100B spent, ~$9.5B recognized back | | Jun 2 | Microsoft | Project Polaris replaces GPT-4 Turbo in Copilot | Default engine from Aug 2026, on Maia silicon | | Jun 5 | DeepSeek | First outside round closes, zero VCs | $7.4B at $52-59B; CATL, Tencent, state funds |
Ten rows, sixty days, one thesis. Control of compute, power, and capital consolidated into a group small enough to fit in one conference room, and two governments now hold seats.
Why this matters if your name isn’t on the table
Here’s the founder translation, and I need you to actually hear it, because the reflex is to file this under “big company news” and get back to your sprint board.
You are not a combatant in the AI infrastructure wars. You are the terrain. Every deal in that table is somebody else making a decision about your unit economics. Your inference costs are set by a pricing committee you’ll never meet, influenced by a subsidy loop you can’t see into. Your model provider’s roadmap bends around its compute landlord. Your “AI margin” is really a bet on how three companies upstream of you decide to split the rent this quarter.
There’s a scene in The Wire where two kids learn chess using street logic, and the lesson that lands hardest is about the pieces that never get to change what they are: the pawns get used early and traded cheap, and “the king stay the king.” The AI stack in 2026 runs on the same rules. The labs, the chipmaker, and the hyperscalers are playing each other. Everyone building on top of them is a piece on somebody else’s board, and no benchmark score changes which side of that line you’re on.
That sounds bleak. It isn’t, quite. Because the same history that produced this consolidation also tells you where the openings are. AWS concentrated computing into three clouds, and the result was the greatest application boom in software history, built by founders who accepted tenancy and priced it in. Stripe concentrated payments and ten thousand companies got built on top. Infrastructure monopolies are terrible things to compete with and wonderful things to build on, as long as you never confuse the two.
Our take: the founders who get hurt in the next 24 months won’t be hurt by AI progress. They’ll be hurt by treating a politically priced input as if it were a stable commodity. And the founders who win will be the ones who treated compute dependency the way good CFOs treat currency exposure: measured, hedged, and reviewed quarterly.
And here’s the part I’m genuinely not settled on. There’s a reading of this whole cycle where the circular money means the demand is partly synthetic, the war is really a bubble inflating its own casus belli, and some of these numbers will look unhinged in three years. I can’t fully dismiss it, and anyone who claims certainty here is selling something. But notice that both readings, war and bubble, give a founder the identical instruction: don’t build your company on the assumption that today’s compute prices, today’s model pricing, or today’s vendor promises survive contact with the next quarter.
What to actually do about it
Five moves, in order of urgency.
First, treat model dependency like currency risk. Microsoft, with the best terms in the industry, still built an exit. You should have one at your scale: a second model wired in behind a feature flag, your prompts and evals portable, your switching cost measured in days. You don’t need to switch. You need the ability to, because the ability is what changes how your landlord treats you.
Second, price the politics into your margin. If your gross margin only works at today’s inference prices, you don’t have a margin, you have a weather forecast. Model a 2x input cost swing in both directions, because both are genuinely possible: DeepSeek’s open-weight pricing drags costs down while compute scarcity drags them up, and the winner of that tug-of-war changes by quarter.
Third, stay off the board you can’t afford to play on. Every few weeks a founder tells me about their plan to offer “cheaper compute” or “independent inference.” Unless you own actual physics, a power source, a cooling trick, real silicon, you are walking into a war between entities that spend your entire lifetime fundraise on a Tuesday. Panthalassa gets to play because it owns waves. A margin-stacking reseller does not.
Fourth, build where the giants’ war makes things cheaper, not scarcer. The application layer above this infrastructure is getting the subsidy of the century: frontier capability at prices held down by competition and state money. The playbook is the same one that always works, which is why we keep pointing founders back to how to validate a startup idea before you bet the runway on it: find a specific customer with a specific pain, and let the warring giants fund your ingredients.
Fifth, watch two dates like a hawk. August 2026, when Polaris becomes Copilot’s default and we learn how sticky an incumbent’s AI distribution really is. And the next Cerebras earnings, when we learn whether an “independent” chip challenger can survive its own customer concentration in public. Those two data points will tell you more about your 2027 costs than any keynote.
The map is the message
When I finally put the string on the wall, the thing that struck me wasn’t the size of the numbers. It was the shape. Every arrow points down the stack. Money, ambition, and state power all flowing away from the glamorous model layer and into dirt, watts, and silicon, the least software-like parts of the software revolution.
The model war had leaderboards, and you could believe your cleverness made you a player. The infrastructure war has landlords, and it doesn’t care how clever you are. It only cares whether you understood, early enough, which one you are.
The giants have made their choice. They’re buying the ground. Your move is simpler and honestly more fun: stop watching their war like a spectator sport, accept that you’re building on leased land, and build the thing the landlords can’t be bothered to build. Tenants outnumber kings. The good ones outlive them too.
Update history: July 7, 2026, first published.
