💻 Computing
TSMC Keeps 8 Cents of Pure Profit From Every Dollar Big Tech Spends on AI. The Price Is Going Up.
We calculated TSMC's extraction rate from the $700 billion AI infrastructure buildout using the company's own earnings data and Big Tech's published capex guidance. The Taiwanese chipmaker captures 15 cents of every dollar and retains 8 cents as net profit, a margin that would make a pharmaceutical company jealous. In 2027, when the price goes up 10 to 15 percent, the surcharge alone will cost AI builders more than OpenAI earns in a year.
TSMC's July 2026 revenue hit NT$467.58 billion. That is $14.50 billion. One month. A 45 percent increase over July 2025, and the company's strongest monthly growth rate this year, accelerating from a January-through-July average of 37 percent. Anthropic, Mistral, and Cohere combined do not generate that much revenue in a full year, yet a single fab operator in Hsinchu pulls it in before August.
Why? Because TSMC fabricates the chips that power artificial intelligence, and it fabricates essentially all of them. Nvidia, AMD, Broadcom, Qualcomm, Apple, every major AI chip designer on Earth sends its designs to TSMC because no one else can manufacture at the required scale and precision. In Q2 2026, high-performance computing accounted for 66 percent of TSMC's $40.20 billion in quarterly revenue, meaning roughly $26.5 billion flowed through its fabs in just three months for AI and HPC chips alone, while advanced technologies at 7 nanometers and below represented 77 percent of wafer revenue, with 3-nanometer at 30 percent and the nascent 2-nanometer node already contributing 3 percent.
The Per-Dollar Extraction Rate
Here is a calculation that, as far as we can find, nobody has published. In H1 2026, TSMC's HPC revenue was approximately $48.4 billion: $21.9 billion in Q1 (61 percent of $35.9 billion) and $26.5 billion in Q2 (66 percent of $40.2 billion). Annualized, adjusted for the accelerating trajectory visible in Q3 guidance of $44.6 to $45.8 billion, TSMC is on pace to generate roughly $105 billion in HPC revenue for the full year. That is just from making other companies' chips.
Big Tech's combined AI capital expenditure for 2026 is approximately $700 billion: Microsoft guided $190 billion, Alphabet guided $180 to $190 billion, Amazon committed $200 billion, Meta ranged $125 to $145 billion, and Oracle set $50 billion. Not all of this goes to chips, of course, because data center construction, power infrastructure, cooling systems, networking equipment, and real estate all count toward the total. But every GPU, every TPU, every custom ASIC in every one of those data centers passes through TSMC's fabrication lines.
Divide TSMC's HPC revenue by Big Tech's AI capex envelope and the extraction rate emerges at 15 cents per dollar, meaning every dollar that Microsoft, Google, Amazon, Meta, and Oracle commit to AI infrastructure sends approximately $0.15 to a single company in southern Taiwan.
Now apply margins, because that is where the number becomes almost absurd. Q2 2026 gross margin hit 67.7 percent, a new record, and net profit margin reached 55.6 percent. At those margins, TSMC retains approximately $0.08 of every AI dollar as pure net profit, which works out to $58 billion in annual AI-derived profit, a figure larger than Luxembourg's GDP, larger than Intel's entire annual revenue, and roughly four times what OpenAI is expected to generate in total revenue this year.
| Metric | Value |
|---|---|
| Big Tech AI capex (2026) | ~$700 billion |
| TSMC HPC revenue (2026E, annualized) | ~$105 billion |
| TSMC share per AI dollar | $0.15 |
| TSMC gross profit per AI dollar (67.7%) | $0.10 |
| TSMC net profit per AI dollar (55.6%) | $0.08 |
| Implied annual AI-derived net profit | ~$58 billion |
Inside the GPU: What TSMC Earns Per Chip
This per-dollar rate becomes concrete inside a single GPU. Take Nvidia's H200, currently the workhorse of AI data centers worldwide, its die fabricated on TSMC's 5-nanometer process. A 300mm wafer at this node costs approximately $20,000, and each wafer yields 6 to 8 usable GPU dies after accounting for defects, putting the bare die cost at $2,500 to $3,300.
But fabrication is only half the bill, because every H200 also requires TSMC's proprietary Chip-on-Wafer-on-Substrate (CoWoS) advanced packaging to bond the GPU die to six stacks of HBM3e high-bandwidth memory on a silicon interposer, a process so complex that it has become as expensive as the chip fabrication itself and so capacity-constrained that Nvidia has locked down 60 percent of all CoWoS expansion through 2027. CoWoS wafer pricing has reached approximately $10,000 per wafer, and each GPU requires roughly one-quarter to one-half of one, adding $2,500 to $5,000 in packaging costs per chip.
Total TSMC revenue per H200 GPU: approximately $5,000 to $8,300, which at 67.7 percent gross margin translates to $3,400 to $5,600 of profit per chip. Nvidia sells each H200 for $30,000 to $35,000, meaning TSMC captures 14 to 24 percent of that final price without ever touching the chip's design, its software stack, or its customer relationship. It prints silicon and packages it, and that turns out to be the most profitable step in the entire AI supply chain.
Why Nobody Competes
Why can nobody undercut TSMC? Because nobody else can manufacture at these nodes. J.P. Morgan analyst Gokul Hariharan estimates that TSMC will retain 95-plus percent market share in the first two waves of 2-nanometer and A16 demand, while Intel's competing 18A node remains in risk production and Samsung Foundry sits a distant third, unable to match TSMC's yields or volume.
Technology alone does not explain the lock-in, though. TSMC's CoWoS advanced packaging capacity is projected to reach 130,000 wafers per month by year-end 2026, potentially scaling to 150,000, and Nvidia has already secured over 60 percent of this capacity for both 2026 and 2027, leaving AMD with approximately 11 percent and everyone else scrambling over the remainder, which means that once a customer commits to TSMC for both front-end wafer fabrication and back-end packaging, switching costs become prohibitive because you are locked in at both stages of the manufacturing process and TSMC sets the price at both. Tape-outs for 2-nanometer designs have already quadrupled the pace of the 3-nanometer cycle, yields are stabilizing between 70 and 80 percent, and nobody is leaving.
The 2027 Price Hike: $23 Billion Added to the AI Bill
TSMC is finalizing negotiations for base price increases of 5 to 10 percent across all nodes, effective early 2027. On top of that, a distinct surcharge: 10 to 15 percent for high-performance computing orders that exceed a customer's initial volume commitments. Every major AI chip customer is capacity-constrained and routinely exceeds initial commitments, putting the effective price increase for AI chips at 12 to 15 percent.
Morgan Stanley projects combined hyperscaler AI capex of $1.2 trillion in 2027 and $1.4 trillion in 2028. If TSMC's extraction rate holds at 15 percent and the effective price increase is 13 percent, the surcharge adds $23 billion to the global AI infrastructure bill in 2027 alone. For perspective: OpenAI's entire 2026 revenue is estimated at $12 to $15 billion. One company's price hike will cost more than the poster child of AI earns in a year.
Who absorbs this? Nvidia and Broadcom, with dominant products and loyal hyperscaler customers, will pass the cost through. AMD, holding roughly 11 percent packaging allocation and competing on price, will likely eat part of it internally. Margins compress. Qualcomm, operating in price-sensitive mobile markets, faces the worst position of all.
The Comparison That Should Unsettle Everyone
TSMC's net profit from AI chip fabrication in 2026 will be approximately $58 billion. Some comparisons:
- OpenAI's estimated 2026 revenue: $12 to $15 billion
- Anthropic's estimated 2026 revenue: $2 to $3 billion
- Combined revenue of every AI-native startup: roughly $25 billion
- Intel's total 2025 revenue: $54 billion
- GDP of Luxembourg: $52 billion
Every AI startup in the world combined earns less in revenue than TSMC earns in profit from making their chips. And the company that profits most from artificial intelligence does not build AI models, does not operate AI services, does not sell AI products. It runs factories. Extremely precise, extremely expensive, extremely difficult-to-replicate factories, sitting on an island 100 miles from a country that has publicly stated its intention to annex it.
The Strongest Counterargument
Here is the strongest case against this framing: TSMC's extraction rate is the natural result of massive capital investment, not monopolistic rent-seeking. Consider the denominator: TSMC is spending $60 to $64 billion in capex in 2026 alone, plus an additional $100 billion committed to Arizona fabs, and those investments must be amortized across the productive life of equipment that will print trillions of transistors before it is retired. A single EUV lithography tool costs $380 million. A single advanced fab costs $20 billion to build. Process R&D runs to billions per node generation. In this context, a 67.7 percent gross margin is not excess profit in the way a software company's margins might be. It is the required return on the most capital-intensive manufacturing on Earth, a return that must fund the next generation of factories before competitors even attempt the current one, because if margins collapsed, TSMC could not afford the jump from 2-nanometer to A14, and the entire AI hardware roadmap would grind to a stop.
Limitations
Several caveats apply. This analysis attributes all of TSMC's HPC segment revenue to AI, which overstates the AI-specific extraction rate because HPC includes non-AI workloads like traditional server CPUs, gaming GPUs, and networking ASICs. TSMC does not disclose the AI-specific fraction. Our $105 billion annual HPC revenue estimate extrapolates from two quarters and assumes Q3 and Q4 follow guidance; demand could soften. Per-GPU cost breakdowns use industry estimates and analyst reports, not TSMC's actual invoices to Nvidia. Big Tech capex includes non-chip spending, so the per-dollar extraction rate represents TSMC's share of the total AI investment envelope, not chip budgets alone. Intel's 18A node could mature faster than expected, though no analyst currently models meaningful share loss for TSMC before 2029.
What You Can Do
If you manage AI infrastructure procurement: request itemized cost breakdowns from your GPU vendor that separate die fabrication, packaging, and markup. CoWoS pricing is climbing faster than wafer pricing. Bundling decisions matter. Choosing packages that use less interposer area can save more than negotiating on chip volume.
If you invest in the AI supply chain: forget both the model layer and the cloud layer, because the highest-margin position in AI is the fabrication layer, held by one company. TSMC's return on equity stands at 40.88 percent, exceeding every major tech company. Its stock carries a geopolitical discount reflecting Taiwan Strait risk, which is real and unhedgeable, but the underlying business extracts rent from every participant in the AI economy without exception. Watch the 2027 price hike, because it is a test of whether anyone will walk away, and nobody will.
If you are building AI products: your cost floor is set in Hsinchu, not on your cloud provider's pricing page. TSMC's margins are the irreducible base cost of every token you generate, every image you render, every model you train. When TSMC raises prices, your unit economics change whether you use Nvidia, AMD, or custom silicon. All roads lead through the same factory.
The Bottom Line
There is a silent tax collector sitting at the center of the $700 billion AI infrastructure buildout. TSMC does not appear on any AI leaderboard. It does not publish benchmark scores. It does not pitch enterprise customers on inference-per-dollar. It just manufactures every chip that every AI company needs, at margins that convert each dollar of industry spending into 8 cents of Taiwanese profit. In 2027, the rate goes up. Its customers have no alternative supplier, no credible second source, and no leverage. The most valuable chokepoint in the AI economy is not a model architecture or a training dataset or a cloud region. It is a set of clean rooms in southern Taiwan that nobody else on Earth can replicate. Everyone on Earth depends on them.