Top AI Infrastructure Stocks 2026: The $1 Trillion Shopping List

Four companies are spending roughly $725 billion this year on AI data centers, and next year they plan to spend a trillion. Amazon, Google, Microsoft, and Meta.^[1]

So, where does all that money go?

It goes to chips. Then memory to feed the chips. Then buildings to hold the chips. Then power to run the buildings. Then somebody to rent it all out. Every one of those stops is a company you can buy.

Of course, the elephant in the room: is this a bubble? Alphabet's stock dropped 7% in July just for raising its capex guide. Meta fell 9% in April for the same sin. Investors are starting to squint at the return on all this spending.

The money for 2026 is already committed. So the real question is 2027-2028, when depreciation from today's spending starts hitting hyperscaler income statements and CFOs get asked hard questions. More on this below.

We've sliced this list by layer of the stack: Silicon (the chips), Memory & Foundry (what the chips are made of and made with), Neoclouds (who rents the chips), Power & Cooling (what keeps them from melting), and the private bellwethers who are writing some of the biggest checks.

This is where the money lands first. When a hyperscaler says "capex," roughly half of it is accelerators. Nvidia owns most of that. Broadcom is the alternative for anyone who wants their own chip design. And Cerebras is the new public kid betting the whole thing shifts to inference.

Nvidia NASDAQ: NVDA

You know Nvidia. Everyone knows Nvidia. But the numbers still make you blink.

Fiscal Q2 2027, reported August 26: revenue of $96.2 billion, up 18% from last quarter and 106% from a year ago. Data center was $89 billion of that. Gross margin held at 75%. And the guide for next quarter is $108 billion, roughly $12 billion more than the record they just set. It was the fourteenth straight quarter of beating their own outlook.

Two things from the report matter for the rest of this list. First, Amazon agreed to buy 2 million Nvidia GPUs and adopt its new Vera CPU. Second, Nvidia's supply commitments more than doubled, from $119 billion to $279 billion, mostly to lock up memory. The biggest chip company on Earth just spent a quarter fighting for DRAM. That tells you where the bottleneck is (see Micron below).

The wrinkle: management said gross margin will drift down to 71-72% by Q4, partly because memory got expensive. Still an absurd number for a hardware company, but the direction changed.

Broadcom NASDAQ: AVGO

If you don't want to pay Nvidia's margin, you design your own chip and hand the blueprint to Broadcom. That's what Google did with the TPU. Meta and OpenAI followed.

The result: Broadcom's AI semiconductor revenue hit $10.8 billion last quarter, up 143%. The guide for the current quarter is $16 billion, up over 200%. Full-year fiscal 2026 AI revenue is expected around $56 billion. And CEO Hock Tan reiterated that fiscal 2027 AI revenue will clear $100 billion, with visibility now stretching into 2028.

Broadcom also sells the networking that stitches a million chips together (its Tomahawk 6 switch went into volume production in March). So even in a world where Nvidia loses share, Broadcom probably wins. That's the appeal: it's the hedge that's also growing 200%.

Cerebras NASDAQ: CBRS

The newest public name on this list. Cerebras makes a chip the size of a dinner plate, one giant wafer instead of a lot of little dies, and claims it runs inference far faster than a GPU. OpenAI already ships a model on it. Amazon's a partner.

It IPO'd on May 14 at $185, opened at $350, raised $5.55 billion, and was the biggest tech listing in years. The stock has since cooled. Revenue was $510 million in 2025 (up 76%), it swung to a profit, and its first public quarter in August showed accelerating sales.

This is the cleanest bet that AI spending shifts from training models to running them. If that happens, speed-per-dollar on inference is the whole game, and Cerebras built for exactly that. The risks: customer concentration, export-control exposure, and a valuation priced for a perfect outcome.

Every accelerator needs high-bandwidth memory bolted to it and a fab to build it. Both are sold out.

Micron NASDAQ: MU

Here's a fun fact: HBM is expected to eat nearly 70% of all DRAM capacity, and the three big memory makers have reportedly already finished allocating their 2027 capacity. Fully booked. Long-term contracts. PCs and phones get whatever's left.

Micron is the American one of those three. Its HBM supply for 2026 is sold out. Its data center segment posted an 87% gross margin on $11.5 billion of revenue last quarter, which is a memory company printing software-like margins. Morgan Stanley called the current shortage the worst in three decades.

The stock's up several hundred percent in a year and dropped about 24% in July even after a record quarter, so it's not calm. But when Nvidia's CFO stands up and says memory scarcity is being caused by the AI buildout itself, and then commits $279 billion to secure supply, you can guess who's on the other side of that check.

TSMC NYSE: TSM

Nvidia designs chips. Broadcom designs chips. Cerebras designs chips. TSMC makes all of them.

Its 3nm lines are running full. Its CoWoS advanced packaging, the step that glues HBM to the accelerator, is quoted at 52-78 week lead times and described as largely booked through 2027. Management's 2026 growth forecast was revised up to 25-30%.

This is the least exciting name here and probably the most durable. If you believe in any of the chip companies above, you believe in TSMC by default, and you get it at a lower multiple. The obvious risk is geographic, and it's not going away.

A neocloud is a company that buys GPUs, plugs them into a data center, and rents them to AI labs that can't get enough capacity from AWS or Azure. Two of them are now real, listed, multi-billion-dollar businesses. They also lose money and carry a lot of debt.

CoreWeave NASDAQ: CRWV

CoreWeave did $2.58 billion of revenue last quarter, up 112%. Its backlog is $104 billion, and that excludes $25 billion of new commitments signed in the first weeks of Q3. Meta added a $21 billion deal on top of a prior $14 billion one. Anthropic signed a multi-year compute agreement. Jane Street committed $6 billion. The stock got added to the Nasdaq-100.

The catch: net loss widened to $626 million, and interest expense alone was $640 million for the quarter. CoreWeave expects to spend $35-39 billion on capex this year. This is a company borrowing enormous sums to buy hardware that depreciates fast, betting the contracts outlast the debt. So far, the contracts keep growing faster than the debt. Watch that ratio every quarter.

Nebius NASDAQ: NBIS

The European challenger. Nebius grew revenue 454% year over year to $582 million, and its AI cloud line grew 514%. Adjusted EBITDA flipped from a loss to $236 million positive. In March it landed a $27 billion contract with Meta and a $2 billion investment from Nvidia.

Smaller than CoreWeave, growing faster, and with a cleaner balance sheet. Michael Burry is short it, which retail traders have been gleefully pointing out as the stock hits new highs. Either way, Nebius is the higher-beta version of the neocloud trade.

The Department of Energy expects data centers to draw 12% of U.S. electricity by 2028. Goldman sees U.S. data center power demand going from 31 GW in 2025 to 66 GW in 2027. Getting electricity to the chips, and getting heat away from them, is now the hard part.

Vertiv NYSE: VRT

Every rack of Nvidia GB300s needs liquid cooling, power distribution, and thermal management. Vertiv sells all three and is working directly with Nvidia on reference designs. It was added to the S&P 500 in March. Analysts expect revenue up about 36% this year and earnings up around 52%.

The risk is lumpiness. Hyperscale projects get built in phases, EMEA sales dipped last quarter, and the stock's beta is over 2. It moves hard both ways on every print. But if the build continues, Vertiv is the most direct public way to own the "physical layer."

GE Vernova NYSE: GEV

Want to know the actual bottleneck in AI? It's gas turbines.

GE Vernova is one of only three companies on Earth that can make a heavy-duty turbine, and it can't make them fast enough. Backlog plus slot reservations hit 116 GW at the end of Q2, up from 83 GW at the start of the year. The company expects at least 125 GW under contract by December and is already taking reservations for 2031 delivery. Power segment orders were up 134% last quarter. Pricing on new orders is running 20%+ above late-2025 levels.

And customers are paying up front for the privilege: Q2 free cash flow hit $5.1 billion, helped by $6.4 billion in deposits. Year-to-date cash generation has already passed all of 2025.

The biggest AI infrastructure spenders after the hyperscalers aren't public. Yet. All three below are either filing or preparing to. And one that was on this list, xAI, went public in June inside SpaceX (NASDAQ: SPCX) after an all-stock merger in February.

OpenAI (Private)

OpenAI is the customer everyone above is building for. Stargate, its $500 billion data center program with SoftBank, Oracle, and MGX, is the single largest compute commitment in the world. It runs on Nvidia GB200s, Oracle operates the sites, and it uses Broadcom and Cerebras silicon too.

OpenAI confirmed on June 8 that it submitted a confidential draft S-1, while saying timing is undecided. CFO Sarah Friar has pointed to late 2026 or 2027. Reported valuation is around $920 billion. Note that Anthropic, the other frontier lab, filed its own draft S-1 on June 1 at a $965 billion valuation.^[2]

Crusoe (Private)

Crusoe built and co-owns the first Stargate campus in Abilene, Texas, financed with roughly $12 billion through 2026 and leased to Oracle. Revenue went from $276 million in 2024 to nearly $1 billion in 2025, with about $2 billion projected this year. Its Wyoming project, "Project Jade," is a planned 2.7 GW campus that would be the largest in the country.

The last confirmed valuation was $13 billion in November 2025. A July 2026 report has it raising at around $30 billion. No S-1 yet, so treat any "pre-IPO Crusoe shares" pitch with heavy suspicion.

Lambda (Private)

Lambda is the developer-friendly GPU cloud, the third-largest behind CoreWeave and Crusoe by revenue. It hired Morgan Stanley, JPMorgan, and Citi to prepare an IPO originally targeted for the first half of 2026. That window came and went without a public filing. With CoreWeave's Q2 reaction and Cerebras's debut, the market's clearly open, so a filing wouldn't surprise us.

Two ways this breaks.

The depreciation wall. The four hyperscalers put $434 billion into property and equipment over the four quarters through March, but their income statements only recognized about $149 billion of depreciation. That gap closes on a five-to-six-year schedule. Meaning 2027-2029 is when today's spending shows up as cost, compressing hyperscaler margins mechanically. The template for what happens next is Amazon in 2022-2023: CFO gets margin questions, capex growth gets moderated, suppliers get whipsawed. Watch for any hyperscaler guiding 2027 capex flat or down, and any extension of server useful-life assumptions (Amazon already reversed one). The July-October earnings cycle is the tell; Alphabet's stock already got punished for raising its guide.

The physical bottleneck. Memory, packaging, turbines, and grid connections all run on a slower clock than chip demand. If gigawatts can't get energized, GPUs sit in boxes and neocloud revenue slips. Nvidia's $279 billion supply commitment is a sign of how tight things are. Watch CoWoS lead times, HBM pricing, CoreWeave's contracted-vs-energized capacity, and GE Vernova's slot conversions.

The next 12-18 months are about three questions.

Does 2027 capex actually hit $1 trillion? Consensus says $935 billion. Guidance comes in January-February with Q4 earnings. If the big four guide up again, every name on this list keeps working. If two of them guide flat, rotate out of the highest-beta names first (Vertiv, Nebius, Cerebras).

Do the mega-IPOs land? OpenAI and Anthropic are both in the pipeline with valuations near $1 trillion each. SpaceX already proved the appetite. Those listings would pull an enormous amount of capital into the theme, and also give the market two very public, very scrutinized numbers for how much a frontier lab spends on compute versus how much it earns.

Does Rubin ship on time? Nvidia's next-gen platform is the engine behind the fiscal 2027 estimates, and CoreWeave has already validated the first Vera Rubin NVL72 system. A clean ramp keeps the Nvidia-Micron-TSMC flywheel spinning. A delay would be the first crack.

Catalysts to circle: Broadcom earnings (early September), Micron earnings (late September), hyperscaler Q3 reports (late October), Nvidia's next print (late November), and any public S-1 from OpenAI or Crusoe.

Own the layers where supply is tightest (memory, packaging, turbines) and don't kid yourself about the leverage in the neoclouds. And keep one eye on that depreciation line.