Nvidia's August 26 Earnings: Why the Revenue Number Isn't the Story

Nvidia reports Q2 fiscal 2027 results after the market closes today.

The number that matters: analysts expect $92.2 billion in revenue — up roughly 13% from Q1's $81.6 billion. The question isn't whether Nvidia will beat that. It's what the rest of the report says about an architectural transition that's happening right underneath the quarterly momentum.

Rubin is entering production now. Inference now accounts for roughly two-thirds of AI compute spend. And Nvidia's biggest customers — Microsoft, Google, Amazon, Meta — are all building their own inference chips.

These three forces converge in one place: this earnings report. Let me lay out the mechanism, because the numbers alone won't tell you whether Nvidia's moat is intact.

Nvidia's last quarter, reported in May, was a masterclass in scale. $81.6 billion in revenue, up 85% from a year earlier. Data Center alone brought in $75.2 billion. Gross margins held at 75%. Operating margins were 64%. The company generated $119 billion in free cash flow over the trailing twelve months while carrying less than $7 billion in net debt. Return on invested capital sits at 89%.

Those are not the metrics of a company at risk. They're the metrics of an infrastructure monopoly — if one that's growing fast enough to justify a $5.16 trillion market cap.

The TTM P/E of 32 looks almost reasonable for that kind of operating quality. But the forward P/E of 57 is what matters for what's coming next. That multiple means the market has already priced in three-plus years of 50% revenue growth. Earnings growth is the only thing supporting it.

Here's the structural shift: AI spend has flipped from training to inference.

During the training phase — where companies build models — Nvidia's CUDA ecosystem was essentially unchallenged. Eighteen years of developer lock-in, tens of millions of lines of optimized code, and a software stack that every AI lab built on top of. Training on anything else meant rewriting your entire framework and accepting unreliable performance.

Inference — running those models at scale for billions of users — cares about different things. Cost per token. Latency. Power efficiency. When you're serving a model continuously to millions of API calls per day, a 30% efficiency gain on alternative silicon pays for the migration in months, not years.

That's why Google built TPU Ironwood, optimized for running Gemini. AWS ships Trainium 3 on its 3nm process. Microsoft's Maia 200 already serves Copilot and OpenAI traffic. Meta's MTIA handles ranking and recommendations.

These are real chips, shipping at hyperscaler scale, and they're all designed specifically for inference workloads.

Here's where the story splits into two markets. Inside the hyperscalers — the companies that serve their own AI products — Nvidia's share is under pressure. Analysts project Nvidia could fall from roughly 90% share on hyperscaler inference to 20–30% by 2028. But that pressure stays inside those four companies. AWS doesn't let you rent Trainium. Google doesn't sell TPU instances outside its ecosystem. Microsoft's Maia is internal only.

For everyone else — the external market, every startup, every enterprise, every AI company that isn't a trillion-dollar cloud provider — Nvidia remains the only practical option. You can't run vLLM, TensorRT-LLM, or FlashAttention on a hyperscaler's custom chip without rewriting the inference stack. The migration costs weeks of engineering time plus ongoing maintenance. For most teams, those costs exceed any savings.

Nvidia does not need a monopoly at 90% on AI accelerators to extend its stock gains — but the inference market will erode the CUDA moat that made that monopoly possible. Slowly, not suddenly. Inside the hyperscalers first, then outward.

This is where Rubin matters.

Nvidia announced at CES this January that Rubin entered full production — eight months ahead of analyst expectations. The development cycle was compressed to 18 months from Blackwell's launch. Volume shipments to cloud partners began in the second half of 2026, which means Q2 is likely the first quarter Rubin starts showing up in revenue.

Rubin is not an incremental upgrade. It delivers up to 10x reduction in inference token cost and 4x fewer GPUs needed to train MoE models, compared with Blackwell. The Rubin GPU carries 336 billion transistors, 288GB of HBM4 memory, and 50 PFLOPS of FP4 inference compute. The Vera Rubin NVL72 rack system — 72 GPUs and 36 CPUs — delivers 3.6 EFLOPS of FP4 inference.

The competitive edge against custom silicon isn't the raw compute. It's the interconnect. NVLink 6 delivers 3.6 TB/s per GPU bandwidth. No custom ASIC has an equivalent. It's the software ecosystem. CUDA doesn't need to be perfect on inference — it just needs to be good enough that the cost of switching still exceeds the savings of custom silicon. Rubin makes the cost of staying with Nvidia substantially lower.

Here's the supply constraint that changes the economics: Rubin production is capped by TSMC's N3 capacity and HBM4 yield rates. Estimated 2026 output is between 200,000 and 300,000 Rubin GPUs. Hyperscalers get priority allocation, likely consuming the first six to nine months of volume. That means Nvidia can sell everything it builds — but it can't build fast enough to saturate demand.

Demand is not the issue. The issue is whether supply commitments, timing, and opportunity cost still justify the same allocation today.

There's a structural concern that doesn't show up in the architecture comparison.

Nvidia is reportedly in discussions to provide approximately $250 billion in financial guarantees to backstop OpenAI's data center leases and chip purchases. At a $5.16 trillion market cap, analysts including Bernstein's Stacy Rasgon describe this as "circular financing" — Nvidia lending money to a customer to buy its own products.

This matters for two reasons. First, it inflates what looks like organic demand. If a significant portion of Data Center revenue is backed by Nvidia's own financial guarantees, the demand signal is weaker than the headline number suggests. Second, it creates contingent liabilities on a balance sheet that currently shows $85 billion in net cash and a debt-to-equity ratio of 4.3%. The leverage isn't there today — but it's there in the commitments.

Bernstein is watching three signals in today's report: whether growth outlook holds, whether gross margins stay in the mid-70s despite rising memory and wafer costs, and whether management provides clarity on those financial guarantees for data center projects valued in the hundreds of billions.

The OpenAI question is the one that doesn't have a clean answer yet. Management commentary today may not resolve it. That's a reason to listen carefully, not a reason to sell.

So what does this earnings report actually tell you?

The revenue number — whether it beats the $92.2 billion consensus — confirms what you already know. Nvidia's scale continues to grow. The sequential growth rate is the more useful metric: if Q2 grows another 13%+ from Q1, it means Rubin is contributing and the ramp is working.

What changes the investment case is management's language on three things. First, Rubin deployment timing and customer qualification — are the hyperscalers actually shipping Rubin systems, or is it still in the pipeline? Second, inference as a revenue driver — Nvidia said the AI chip opportunity could reach $1 trillion through 2027. That number only works if inference adoption keeps accelerating. Third, supply visibility — can Nvidia see strong demand through the end of the fiscal year, or is there any hint of softening?

The valuation question is separate from the earnings result. At a forward P/E of 57, the stock already assumes flawless execution through 2028 and beyond. That doesn't mean the valuation is wrong — earnings growing at 50%+ can compress a 57 multiple into the 30s within two years even if the stock price stays flat. But it means there's no margin for error on guidance.

While I still believe Nvidia's architecture cycle gives it years of compounding ahead, the return profile is no longer as straightforward as it was at a forward P/E of 30. Much of the total return is likely to be back-half weighted in 2028-2030, as Rubin scales, inference becomes a larger revenue share, and the custom silicon threat resolves itself in one direction or the other.

The debate is not about whether Nvidia stays important. It is about whether the return profile is still as compelling as what can be found elsewhere in the AI trade.

Today's report won't answer that fully. But the numbers, the margins, and Jensen Huang's language on Rubin, inference, and the supply chain will tell you whether the thesis that got you here still holds — or whether it needs to be adjusted for where the company actually is.