Nvidia Q2 Revenue Hits $96.2B Despite Rising Custom TPU Pressure

A blowout quarter in a crowded market

Nvidia's second quarter of fiscal 2027 beat Wall Street's estimates on almost every line. This happened while the hyperscalers that buy its chips were building more of their own. For the quarter ended July 26, 2026, the company reported revenue of $96.2 billion. That was up 18% from the prior quarter and 106% from a year earlier. Gross margin was 75.0% on both a GAAP and a non-GAAP basis.11 Data center sales brought in $89.0 billion, up 117% year over year.15 GAAP diluted earnings came to $2.46 per share and non-GAAP earnings to $2.22.19

Against consensus, revenue beat the analyst average of about $92.07 billion by roughly 4.5%, and non-GAAP EPS beat the $2.09 forecast by about 6%.12 The company also guided third-quarter revenue to $108.0 billion, plus or minus 2%. That guidance assumes no data center compute revenue from China.11 It was about 4% above the roughly $103.8 billion the Street expected.12 By one count, this was the fourteenth quarter in a row that Nvidia beat its own outlook.12

Nvidia remains the main beneficiary of AI infrastructure spending. Still, much of the year's coverage has asked whether custom silicon, led by Google's Tensor Processing Units, will eventually cut into that position. The quarter offers a partial answer: not yet, and not in a way that shows up in Nvidia's revenue.

Where the demand came from

The customer breakdown matters more than the headline total. On the earnings call, management said hyperscale cloud providers accounted for about $49 billion of data center revenue, up 13% sequentially.20 The remaining $40 billion came from a category Nvidia calls ACIE. It covers AI-native companies, "neoclouds," enterprises and sovereign programs. That business grew 25% sequentially and 138% year over year.20 CFO Colette Kress said it now makes up roughly half of the data center business.14 Sovereign AI revenue reportedly tripled from a year earlier.16

This mix explains why the custom-silicon threat has not yet reached the income statement. Google, Amazon, Microsoft and Meta can design their own accelerators. Most governments, enterprises and smaller GPU clouds cannot. Nvidia's fastest-growing customers are the ones least able to build alternatives. A growing share of its revenue comes from buyers who are hard to win over with in-house chips.

Nvidia is also earning more from each unit of data center capacity. Management put its revenue opportunity at about $18 billion per gigawatt with Hopper and $25 billion with Grace Blackwell. It expects around $40 billion with Vera Rubin, counting CPUs, networking and Groq technology in the bundle.20 The company also said Vera Rubin is ramping into full production, with racks already running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius.11 On the call, Nvidia highlighted a larger partnership with AWS covering roughly 2 million more GPUs.20

The TPU challenge is real

Google's custom silicon is no longer a side project. In April, Google Cloud announced its eighth-generation TPU and split it into two chips: the TPU 8t for training and the TPU 8i for inference.1 Google claimed up to 3x faster training, 80% better performance per dollar, and clusters of more than a million TPUs.1 Yahoo Finance reported that the TPU 8t offers 2.8x better price-performance than the previous generation.2 Amin Vahdat, Google's chief technologist for AI infrastructure, said the TPU 8i pairs high-bandwidth memory with three times as much on-chip SRAM to reduce latency between AI agents.10

The customer list now includes the biggest names in AI. Google expanded its deal with Anthropic to provide "multiple gigawatts" of next-generation TPU capacity. It is also working to supply TPU capacity to OpenAI. Meta reportedly signed a multiyear, multibillion-dollar TPU agreement.2 One analysis described the Anthropic expansion as up to 5 gigawatts of capacity, tied to a Google investment of up to $40 billion in the lab.3 Google is also moving TPUs beyond its own cloud. It reportedly provided a $3.2 billion guarantee for a New York data center that will lease TPU capacity to Anthropic, and it confirmed plans to sell TPUs directly to outside buyers.7 A Blackstone joint venture aims to deploy 500 megawatts of TPU capacity by 2027.8

Customers say the cost savings are real. Citadel Securities reported that moving research workloads to TPUs ran them up to four times faster at about 30% lower cost.7 Earlier industry reporting cited one company that cut monthly compute costs by about 65% after moving inference from Nvidia GPUs to TPUs.5 Some analysts expect Nvidia's share of inference to fall from about 90% to 20–30% by 2028 as hyperscalers move internal workloads onto their own chips.4 Inference already accounts for roughly two-thirds of AI compute spending.4

Where the coverage diverges

Outlets do not agree on how much ground Nvidia has lost. Estimates of its market share include 92% of the data center GPU market10, about 86% of AI data center compute8, roughly 90%7, and 81% of the broader AI chip market.3 The differences come mostly from definitions. A GPU-only market leaves out TPUs entirely, while an "AI chip" market includes them. The direction is the same in every case: Nvidia is still dominant, and its share is shrinking at the edges.

The timelines also differ. TechCrunch and Yahoo Finance dated Google's TPU 8 launch to Cloud Next in April 2026.12 Another roundup put the launch in August and said Rubin's general availability was targeted for the first quarter of 2027.6 Nvidia's own earnings release says Rubin racks are already running at five major clouds.11 The April date has the stronger sourcing, and Nvidia's statement about Rubin production is a direct company disclosure. The later dates look like errors in secondary coverage.

The biggest disagreement is about what custom silicon means for Nvidia's margins. The Motley Fool argued that TPU competition could erode Nvidia's pricing power and its gross margin of about 74%.8 The latest quarter does not show it. Margin held at 75.0%, up about 2.5 points from a year earlier, and guidance calls for about 74.0% next quarter.1511 Another analysis flagged rising DRAM costs as an emerging pressure on margins.16 For now, memory costs look like a bigger near-term risk to margins than Google's chips.

Why custom silicon hasn't dented Nvidia yet

The most useful detail in the TPU coverage may be that Google is one of Nvidia's customers. Google Cloud supplements its Nvidia systems with TPUs rather than replacing them, and it committed to offering Vera Rubin.1 Alphabet still depends heavily on Nvidia GPUs and is among the company's largest buyers.10 Anthropic added Google as a supplier without dropping anyone. It runs workloads across TPUs, AWS Trainium and Nvidia GPUs.3 Jensen Huang has publicly questioned whether custom ASICs really cost less than general-purpose GPUs.7

The likely explanation is that demand is outrunning every supplier. Coverage of Google's push noted that no single vendor can comfortably meet current compute demand.7 Nvidia itself described the quarter as supply-constrained.20 In that market, TPUs gain share by adding capacity that Nvidia could not have delivered anyway. They are not yet taking existing sales away from it. TSMC builds chips for Nvidia, Google's partner Broadcom, and nearly everyone else, which makes it the safest winner of the contest.3

What to watch

The risk for Nvidia is the rate of growth, not the level of demand. Its guidance implies a sequential increase of about $11.8 billion. That is smaller than the roughly $14.6 billion jump from the first to the second quarter, so the dollars added each quarter are no longer increasing.12 When supply catches up with demand, buyers will be able to choose more freely between GPUs and alternatives. At that point, the cost advantages Google and its customers report will start to affect pricing.

Until then, the quarter supports one conclusion. Custom silicon is a serious long-term threat to Nvidia's share of hyperscaler inference. In the near term, it is mostly helping to expand total AI compute capacity. Nvidia's results show the pie is still growing faster than rivals can take slices, and its fastest growth is coming from customers who cannot build their own chips.