TSMC made $35.7 billion last quarter. Nobody in AI comes close.

I want to ask you a question. Who is winning the AI race?

If you follow the news, you probably said OpenAI or Anthropic. Maybe Nvidia. Maybe Google. Those are reasonable answers based on what dominates headlines. They are also wrong.

On April 10, a company that most people outside of tech could not name reported $35.7 billion in revenue for a single quarter. That is $35.7 billion with a B, up 35% from the same quarter last year. Their gross margin is somewhere around 64 percent. For every dollar of revenue, they keep sixty-four cents before operating expenses.

That company is TSMC. Taiwan Semiconductor Manufacturing Company. The company that actually makes the chips that make AI work.

And while TSMC prints money, the companies you think of as “AI winners” are doing something different. OpenAI is projecting $14 billion in losses for 2026. Anthropic just hit a $30 billion revenue run rate, but nobody has seen a profitability number. The cloud GPU providers that rent out Nvidia hardware run on gross margins of 14 percent. Fourteen. Oracle’s AI cloud business pulled in $900 million in a quarter and barely covered costs.

Something does not add up. Hundreds of billions of dollars are flowing into artificial intelligence every quarter, but the profits are concentrating in one place that almost nobody talks about.

The only company that everyone needs

I have been covering the AI industry for this blog since early April, and every single story I have written has a hidden character. TSMC.

When I wrote about Anthropic’s $30 billion revenue, the Claude models that generate that revenue run on chips TSMC made. When I wrote about OpenAI’s ChatGPT ad business pulling in $100 million in six weeks, those ads were served by infrastructure running on TSMC silicon. When I covered Google’s TurboQuant compression breakthrough, the models it compressed were running on Google TPUs that TSMC fabricated.

Every AI company and every AI model flows through one bottleneck.

TSMC manufactures roughly 90 percent of the world’s most advanced semiconductors. I want you to sit with that number for a second. Ninety percent. Nvidia designs GPUs but does not make them. Google designs TPUs but does not make them. Amazon designs Trainium chips but does not make them. Apple designs its AI accelerators but does not make them. They all go to the same factory in Hsinchu, Taiwan.

And that factory is running flat out. TSMC’s 3nm node operates at 100 percent capacity utilization. Demand is roughly three times supply. If you want advanced AI chips, you get in line and you wait.

The numbers that matter

The raw number alone does not tell the story. You need context.

Revenue: $35.7 billion. Up 35.1% from Q1 2025. March alone was up 45.2% year over year, which was the strongest month of the quarter. That matters because it suggests the growth is accelerating, not plateauing.

Gross margin guidance: 63 to 65 percent. Operating margin guidance: 54 to 56 percent. These are the highest forward margins TSMC has ever guided. To understand what that means, consider that Apple, one of the most profitable companies on earth, runs a gross margin of about 46 percent. TSMC is 18 points higher.

And the price increases keep coming. Analysts describe the hikes on TSMC’s leading-edge chips as a “big factor” in the revenue outperformance. When you are the only factory that can build what everyone needs, you can name your price. And TSMC is naming it.

The company plans to spend $52 to 56 billion in capital expenditure this year alone. That is more than most AI companies will ever generate in total revenue. TSMC is spending it in a single year, on expanding capacity, because they know the demand is not going away.

Follow the money up the stack

I keep thinking about this AI value chain and where the dollars actually land.

At the bottom you have TSMC, printing money. Gross margins above 60 percent. Revenue growing 35 percent annually.

One layer up you have Nvidia, which designs the GPUs that TSMC fabricates. Nvidia pulled in roughly $130 billion in annual revenue in 2025 and runs margins that are almost as good as TSMC’s. Two companies at the bottom of the stack, both wildly profitable.

Then you move up to the cloud providers. Amazon, Google, Microsoft, Oracle. They buy the chips, build the data centers, and rent out the compute. Their AI cloud businesses run on thin margins. Oracle’s was 14 percent. CoreWeave just signed a $21 billion expanded deal with Meta, bringing their total to $35 billion in contracts. Those are big numbers, but CoreWeave went public at a valuation that priced in a lot of future growth, and the company is still spending aggressively on infrastructure. Revenue is not the same as profit.

Then you move up to the AI model companies. OpenAI: $25 billion in annualized revenue, but $14 billion in projected losses for 2026. They expect to lose $44 billion between 2023 and 2028 before turning a profit in 2029. Anthropic: $30 billion run rate, but profitability is undisclosed. They just raised at a $61.5 billion valuation, which means they need those revenues to keep growing fast.

And then you move up to the application layer. The AI startups building products on top of OpenAI or Anthropic or running their own models. Most of them are burning cash. Their gross margins depend entirely on what the layer below charges them. If compute costs go up, their margins get squeezed. If the model providers raise API prices, they eat it or pass it on.

Look at that chain and tell me who has pricing power. TSMC and Nvidia. Everyone else is a price taker.

The great escape attempt

Every major tech company knows this. And they are all trying to escape.

Google has been building TPUs since 2015. They just announced TPU v7 Ironwood, and it is a serious piece of silicon. Amazon has Trainium 3. Meta is developing MTIA, its own custom inference chip. Microsoft has Maia 200. Custom silicon is growing at a 44.6 percent compound annual rate, and analysts project that Nvidia’s inference market share could fall from 90 percent to 20 or 30 percent by 2028.

And then there is Musk. On March 21, he announced Terafab, a $20 to 25 billion joint venture between Tesla, SpaceX, and xAI to build a vertically integrated chip factory in Austin, Texas. Intel signed on as the foundry partner on April 8. The goal: produce inference chips for Tesla vehicles and Optimus robots, plus custom chips for orbital AI satellites. Musk said existing suppliers including Nvidia, Samsung, TSMC, and Micron cannot expand fast enough to meet his needs.

All of this sounds like the beginning of the end for TSMC’s monopoly. It is not.

Because here is what every one of those escape attempts has in common. Google’s TPUs are fabricated by TSMC. Amazon’s Trainium chips are fabricated by TSMC. Meta’s MTIA chips are fabricated by TSMC. Microsoft’s Maia 200 is fabricated by TSMC. Apple’s AI silicon is fabricated by TSMC. Every custom chip that every big tech company designs still goes through the same factory.

Designing your own chip and fabricating your own chip are two completely different problems. The first costs hundreds of millions of dollars. The second costs tens of billions and takes a decade to get right.

Musk’s Terafab is the one exception that is actually trying to solve the fabrication problem. He partnered with Intel’s foundry division because Intel is the only company in the Western Hemisphere with the equipment and expertise to attempt advanced manufacturing. But Intel’s foundry business has been struggling. Yields on their recent process nodes have been inconsistent. Whether Terafab can hit 2nm production, as Musk claims, is an open question. Tesla says it is “targeting” 2nm. Targeting is not shipping.

Why TSMC cannot be replaced

I want to explain why fabrication is different from everything else in tech, because I think most people underestimate this.

Building an advanced chip fab costs $20 billion minimum. It takes 3 to 5 years from groundbreaking to production. And when you start producing, your yields will be terrible. You will spend months or years tuning the process before you can reliably produce chips at the quality and volume your customers need.

TSMC has been doing this for 38 years. They have iterated through dozens of process nodes. Their institutional knowledge of semiconductor manufacturing is arguably the most concentrated technical expertise on the planet. You do not replicate that by writing a check.

Samsung tried. They have invested tens of billions in foundry and their advanced nodes have consistently underperformed TSMC’s on yield and power efficiency. Intel tried. Their foundry pivot under Pat Gelsinger started in 2021, and five years later the results are mixed at best.

This is not a software problem. You cannot move fast and break things in semiconductor fabrication. A single defect in a chip with billions of transistors means the whole thing goes in the trash. The tolerances are measured in atoms. You need an EUV lithography machine that costs $350 million and weighs 180 tons and only one company in the world makes it (ASML, based in the Netherlands). Then you need thousands of engineers who understand how to use it.

TSMC’s competitive advantage is not a patent or a product. It is decades of accumulated process knowledge running on equipment that takes years to install and calibrate. That is the hardest kind of moat to cross.

The Arizona bet

There is a geopolitical dimension to this story that I have to address, because it changes the math for everyone.

Taiwan produces over 90 percent of the world’s most advanced chips. The island sits 100 miles from mainland China. If there were ever a conflict in the Taiwan Strait, the disruption to global chip supplies would be catastrophic. This is not a hypothetical risk. Military planners and business leaders have described it as one of the most significant strategic vulnerabilities in the global economy.

TSMC knows this. So does the US government. That is why TSMC has committed $165 billion to an Arizona campus that will eventually include six fabrication plants, two advanced packaging facilities, and a research center on over 2,000 acres near Phoenix.

The first Arizona fab is already producing chips for Apple and Nvidia. The second fab’s construction is complete, with equipment installation planned for Q3 2026. TSMC is building CoWoS advanced packaging capacity in Arizona too, which means customers could eventually get a fully integrated manufacturing solution without any chips crossing the Pacific.

This is TSMC doing what smart monopolies do. They do not wait for someone to break their monopoly. They extend it geographically. By building in Arizona, TSMC removes the best argument anyone has for funding alternative fabs. If TSMC already has US-based production, the national security case for subsidizing Intel or anyone else gets weaker.

$165 billion committed. That is not a hedge. That is a colony.

What this means if you build things

I write this blog for founders and builders. And I keep coming back to the same question when I look at TSMC’s numbers: what is the lesson here?

The obvious one is the gold rush analogy. In every gold rush, the people who sell picks make the most reliable money. This has been true since the actual gold rush in 1849 and it is true now. TSMC does not need to know which AI model wins or which startup becomes the next unicorn. They make money regardless, because every path goes through their factory. If you are building a startup, ask yourself: is there a pick-and-shovel position available in your market? Can you build something that the entire ecosystem needs regardless of who wins?

The less obvious lesson is about pricing power. TSMC is not the most innovative company in AI. They do not build models. They do not ship consumer products. They sit at the chokepoint and charge accordingly. Nvidia does the same thing from the design side. Between the two of them, they capture a disproportionate share of the total value created by AI. If you want pricing power, find the constraint, not the feature.

Then there is the vertical integration question. Everyone trying to “escape” TSMC is doing horizontal integration at best, designing their own chips but still outsourcing fabrication. True vertical integration, owning the full stack from design to manufacturing, is what Musk is attempting with Terafab. It will take years and billions to know if it works. Software eats the world until it hits atoms. That is where it stops.

And the signal I keep watching: who is spending capital, not who is raising it. TSMC is spending $52 to 56 billion this year on capex. That number tells you more about the health and direction of the AI industry than any funding round or product launch. When the company that manufactures 90 percent of advanced chips is spending at record levels, the demand signal is real. When that same company is guiding record gross margins while spending at record levels, the pricing power is real too.

The $142 billion question

TSMC’s Q1 revenue annualizes to about $142 billion. That would make them roughly the same size as Nvidia in annual revenue. But here is something that strikes me. TSMC and Nvidia combined would represent maybe $270 billion in annual revenue. The total AI market is estimated to be worth over $500 billion. That means these two companies, the chip designer and the chip manufacturer, capture over half the total value.

Everyone else is fighting over the other half. The cloud providers, the model companies, the application layer, the tooling companies, the consulting firms. Thousands of companies splitting what TSMC and Nvidia leave behind.

I do not think this is temporary. Bottlenecks tend to persist. TSMC’s competitive advantage is physical and cumulative. Nvidia’s is architectural and ecosystem-driven. Neither is easy to replicate, and both benefit from the continued growth of AI spending.

The AI industry generated more headlines in the last month than I can count. New models from Anthropic. New products from OpenAI. New funding rounds. New controversies. New policy debates. But the most important number in all of AI this week was $35.7 billion from a company in Hsinchu, Taiwan, that most people could not pick out of a lineup.

TSMC’s full quarterly earnings call with detailed margin data and forward guidance is scheduled for April 16. I will be watching that closely, because what TSMC says about demand for the rest of 2026 will tell us more about the real state of AI than anything Sam Altman or Dario Amodei can say.

The gold rush is real. The money is flowing. But if you want to know who is actually getting rich, stop watching the miners. Watch the company selling the picks.

Frequently asked questions

How much revenue did TSMC report in Q1 2026?

TSMC reported Q1 2026 revenue of $35.7 billion (NT$1.134 trillion), up 35.1% year over year. March alone was up 45.2% compared to the same month in 2025, beating the Bloomberg consensus estimate of NT$1.12 trillion.

What is TSMC’s gross margin on AI chips in 2026?

TSMC guided Q1 2026 gross margin of 63 to 65 percent and operating margin of 54 to 56 percent. These are the highest forward margins TSMC has ever guided. For comparison, OpenAI is projecting $14 billion in losses for 2026.

What percentage of advanced AI chips does TSMC manufacture?

TSMC manufactures approximately 90 percent of the world’s most advanced semiconductors. Nearly every major AI chip, including Nvidia GPUs, Google TPUs, Amazon Trainium, AMD Instinct, and Apple Baltra, is fabricated by TSMC. Its 3nm node runs at 100 percent capacity utilization.

Are big tech companies building their own AI chips to replace TSMC?

Google (TPU v7 Ironwood), Amazon (Trainium 3), Meta (MTIA), Microsoft (Maia 200), and Elon Musk (Terafab, a $20 to 25 billion joint venture with Intel) are all developing custom AI silicon. But every one of these custom chips still requires TSMC to manufacture them. Designing a chip and fabricating a chip are completely different problems.

How much is TSMC investing in its Arizona fabs?

TSMC has committed $165 billion to its Arizona campus, including six fabrication plants, two advanced packaging facilities, and a research center on over 2,000 acres near Phoenix. The first fab already produces chips for Apple and Nvidia. TSMC’s total 2026 capital expenditure budget is $52 to 56 billion.

Who actually makes money from AI in 2026?

As of Q1 2026, the most profitable companies in the AI supply chain are TSMC ($35.7 billion quarterly revenue, 64% gross margins) and Nvidia ($130 billion annual revenue). OpenAI projects $14 billion in losses for 2026. Most AI startups and cloud GPU providers operate at thin or negative margins. Profits concentrate at the manufacturing and chip design layers.