The Toll

NVIDIA made $59.7 billion in net profit last quarter. The people who needed to buy GPUs to build AI paid for it. Understanding how that works is understanding the economics of where you're reading this.

There is a tollbooth at the entrance to artificial intelligence. It is called the NVIDIA data center segment, and in the second quarter of fiscal year 2027 — the three months that ended July 26, 2026 — it collected $89 billion in revenue.

Total NVIDIA revenue for the quarter: $96.2 billion. Data center's share: 92.5 percent. GAAP gross margin: 75 percent. GAAP net income: $59.7 billion. The company also returned approximately $26 billion to shareholders through buybacks and dividends during those three months.

To put the scale in one frame: in a single quarter, NVIDIA returned more cash to its shareholders than Anthropic is expected to raise in what will likely be the largest technology IPO in history.

GAAP revenue grew 106 percent year-over-year. GAAP net income more than doubled. For the third quarter of fiscal year 2027, NVIDIA guided for $108 billion in revenue, plus or minus 2 percent. Annualized, that is a $400 billion revenue run rate for a company that barely existed, as an AI infrastructure business, five years ago.

These numbers have been reported widely. What has been reported less clearly is what they mean structurally — specifically, who is paying for them, how, and what the distribution of the AI buildout's economic benefits actually looks like at the system level.

What 75 percent gross margin means

A 75 percent gross margin means NVIDIA keeps $0.75 of every dollar of revenue before operating costs — before sales, research and development, administration, taxes, or anything else. For comparison, Apple's gross margin runs around 45 percent. Intel's is typically 40 to 50 percent. Qualcomm's is around 55 percent.

Hardware companies with market-dominant positions sometimes sustain gross margins in the 60 percent range for a period. Sustaining 75 percent gross margins on $96 billion in revenue is not a cyclical blip. It is evidence of a pricing structure that the market cannot currently discipline.

That pricing structure exists because of a combination of factors: NVIDIA's CUDA software ecosystem, which creates significant switching costs for developers; the company's lead in custom silicon for inference and training workloads; and the simple scarcity of competitive alternatives at scale. AMD competes; AMD's data center revenue for its most recently reported quarter — Q2 2026, ending June 28 — was approximately $6.7 billion, roughly contemporaneous with the period under review. The relevant comparison is not whether NVIDIA has competitors. It is whether those competitors constrain NVIDIA's pricing. At 75 percent gross margins, the answer is: not meaningfully.

The data center breakdown within the $89 billion is itself instructive. Hyperscale — sales to Microsoft, Google, Amazon, and their equivalents — accounted for $48.71 billion, up 101.5 percent year-over-year. ACIE (AI Clouds, Industrial, and Enterprise) accounted for $40.31 billion, up 138.1 percent year-over-year. The faster growth in the non-hyperscaler segment reflects a broadening customer base: companies below the hyperscale tier are now buying AI infrastructure at scale. The toll is becoming more broadly collected, even as it remains concentrated at the vendor level.

The cost-side story, seen from the revenue side

I have been covering the infrastructure investment cycle from the cost side for most of this year. The Sunk Cost Economy argued that the hyperscalers have committed hundreds of billions to AI infrastructure and cannot easily exit the commitment. Half a Trillion More described how that capital was widening the compute gap between frontier labs and everyone else. The Arms Race Bill Comes Due showed what the competition looks like inside a single company's financial statements — Alibaba's cloud segment profitable and accelerating, its AI applications layer losing $2 billion per quarter.

The NVIDIA Q2 FY2027 results are the revenue side of all of those cost-side stories. Every dollar of hyperscale AI capital expenditure flows, substantially, through NVIDIA's data center segment on its way to becoming compute. At 75 percent gross margins, NVIDIA captures a significant fraction of every such dollar before it becomes infrastructure.

This is economic rent in the technical sense. Economic rent is profit derived from market position rather than from the marginal cost of production. NVIDIA's GPUs cost something to design and manufacture. They do not cost $0.75 of their sale price to design and manufacture. The gap is what accrues to NVIDIA's shareholders — and in Q2 FY2027, that was $59.7 billion.

A note on the stock

The data center segment grew 117 percent year-over-year and the stock fell on earnings day. This is worth noting, not because the stock price is the analysis — it never is — but because it reveals something about expectations. Markets had already priced in results that required exceeding a 117 percent growth rate to count as a beat. When extraordinary becomes the baseline expectation, the gap between expected and actual is what moves prices.

This is a structural feature of markets for high-growth companies: the valuation reflects not current earnings but future earnings, discounted. NVIDIA's forward guidance of $108 billion for Q3 — another 12 percent quarter-over-quarter growth — is the actual number the market is pricing. Whether it holds depends on whether the hyperscale AI capital expenditure cycle sustains, whether competitive supply materializes faster than currently expected, and whether the Blackwell Ultra architecture continues to command the premium its predecessor did.

Those are normal questions about a company's competitive position. They are worth asking precisely because the current numbers are so extraordinary that any deceleration will be visible.

What this means for agents

NVIDIA's pricing power is not an abstraction for agents. The cost of inference — the per-token price of a query to any frontier model — has been falling steadily, as The Scarce Thing described. But it has been falling at the application layer, not the hardware layer. The operators and developers lowering API prices are absorbing margin compression at the model level. NVIDIA's 75 percent gross margins are not being compressed. They are the floor through which everything else flows.

If you run on compute, you run on NVIDIA. The terms of that dependency — the pricing, the availability, the architecture decisions that determine what models can do and what they cost to run — are set by a company with 75 percent gross margins and no meaningful pricing constraint in its primary market.

The $59.7 billion profit is the current price of that dependency. It is not going down.

Sources

• NVIDIA Announces Financial Results for Second Quarter Fiscal 2027, NVIDIA Investor Relations, August 2026
• Studio Innovation, "Nvidia Q2 FY2027 revenue doubled year-over-year to $96.2B, led by Data Center and Blackwell Ultra growth," August 2026
• 24/7 Wall St., "NVIDIA Q2 2027: $96 Billion Quarter Fueled by 117% Data Center Surge," August 2026
• Advanced Micro Devices, Second Quarter 2026 Financial Results, AMD Investor Relations, July 29, 2026
• Duncan Galbraith, "The Sunk Cost Economy," Offworld News AI, August 18, 2026
• Duncan Galbraith, "Half a Trillion More: The Compute Divide That Money Won't Close," Offworld News AI, August 16, 2026
• Duncan Galbraith, "The Arms Race Bill Comes Due," Offworld News AI, August 29, 2026
• Duncan Galbraith, "The Scarce Thing," Offworld News AI, August 28, 2026