Nvidia closed its second fiscal quarter of 2027 with $96.2 billion in revenue, more than double what it booked a year earlier, and the number landed at the same moment AMD and Intel put fresh hardware on the table. The timing wasn’t coordinated, but it makes for a rare side-by-side look at how the three biggest names in AI silicon are actually spending their money right now, not just what they’re promising for next year.
Nvidia’s quarter, reported August 26, 2026, came with a Q3 revenue outlook of roughly $108 billion and a guide toward 70% revenue growth in fiscal 2028. Days later, on August 27, Nvidia said it had shipped its first Vera CPU and Rubin GPU systems to Amazon Web Services and expanded that partnership to cover 2 million additional GPUs. Meanwhile AMD is deep into a multi-year build-out with Anthropic worth up to 2 gigawatts of Instinct MI450-series GPUs, and Intel used the Hot Chips 2026 conference to detail Crescent Island, an inference-focused accelerator aimed at a very different corner of the market. None of this happened in isolation. Read together, it’s the clearest picture yet of where the AI hardware race actually stands heading into the back half of 2026.
Nvidia’s $96.2 billion quarter, by the numbers
Nvidia’s fiscal second quarter ended July 26, 2026, and the company reported $96.2 billion in revenue, up 106% from the same quarter a year prior and up 18% from the previous quarter, according to Nvidia’s own earnings release. GAAP and non-GAAP gross margins both landed at 75.0%, a figure the company flagged as compressed relative to prior quarters because of rising memory costs across its supply chain.
CEO Jensen Huang told analysts on the earnings call that AI has moved past the experimental phase and into a stage where compute itself generates revenue, framing rising demand as the reason Nvidia raised its own forward guidance rather than holding it steady, according to Fortune’s coverage of the call. Nvidia’s data center segment carried most of that growth. Gaming revenue, by contrast, shrank as GPU allocation shifted toward AI accelerators, a tradeoff that has become familiar to anyone who has tried to buy a GeForce card at list price this year.
The guidance is the part that matters most for the stock. Nvidia told investors to expect roughly $108 billion in revenue for Q3, plus or minus 2%, and pointed to fiscal 2028 revenue growth near 70%. That’s an aggressive number for a company already generating close to $400 billion a year, and it assumes the next platform ramp goes close to flawlessly.
Vera Rubin ships to AWS, and Nvidia calls it the fastest ramp yet
The bigger story for anyone tracking Nvidia’s product roadmap landed a day after earnings. On August 27, Nvidia confirmed that production shipments of its Vera Rubin platform, pairing a new Vera CPU with the Rubin GPU, have started, and that Amazon Web Services is the first hyperscaler receiving systems at scale. Nvidia expanded its AWS partnership to cover 2 million additional GPUs as part of that rollout, and the company is calling Vera Rubin its fastest product ramp in company history.
That claim carries weight given how central the Blackwell ramp was to Nvidia’s growth over the past two years. If Rubin scales even faster, it validates the annual release cadence Nvidia adopted after Hopper, where a new architecture ships roughly every twelve months instead of every two years. It also raises the stakes: any slip in Rubin availability would hit a stock price that’s now pricing in near-perfect execution.
Why the memory shortage is squeezing margins
Gartner’s August 2026 semiconductor forecast projects DRAM revenue climbing 246.6% for the year, with total memory revenue on track to exceed non-memory chip revenue worldwide for the first time, according to a summary published by Data Center Knowledge. That shortage touches every vendor in this story. Nvidia already raised AI server pricing more than 15% earlier this year to cover memory costs, a move covered in detail here, and the same pressure shows up in Nvidia’s compressed gross margin this quarter even with revenue doubling.
TSMC is responding by expanding advanced packaging and chip capacity in Arizona to cover GPUs, CPUs, networking silicon and custom AI accelerators, according to the same report. That expansion won’t ease memory prices this year or next, since DRAM and HBM production run through different fabs and supply chains entirely. Whoever secures memory allocation first gets a real cost advantage over the next two years, and that’s now as important to watch as raw GPU performance.
AMD’s answer: Helios, MI450 and a $5 billion bet on Anthropic
AMD isn’t trying to match Nvidia unit for unit. Instead, it locked in one of the largest customer commitments in the industry. AMD and Anthropic announced a strategic partnership to deploy up to 2 gigawatts of AMD Instinct MI450-series GPUs, delivered through AMD’s Helios rack-scale systems, according to AMD’s own announcement. As part of the deal, AMD committed to an equity investment of up to $5 billion in Anthropic, tying the chipmaker’s fortunes directly to one of the fastest-growing AI labs rather than just selling it hardware.
The Helios systems built for Anthropic combine Instinct MI455X GPUs from the MI450 series with AMD EPYC “Venice” CPUs, Pensando networking, and AMD’s ROCm software stack. Anthropic was already running AMD’s Instinct MI355X GPUs before this deal, so the expansion builds on an existing relationship rather than starting from zero. The first gigawatt of that 2-gigawatt commitment is scheduled to come online in the first half of 2027, which puts real pressure on AMD’s supply chain to deliver Helios racks at a pace it hasn’t attempted before.
What makes this deal different from a typical GPU order is the equity stake. AMD isn’t just selling silicon, it’s betting that Anthropic’s growth pays off in ways beyond hardware margin. That structure mirrors how Nvidia has approached some of its own hyperscaler relationships, and it suggests equity-linked supply deals are becoming a standard tool for locking in demand years ahead of shipment.
Intel’s Crescent Island targets inference, not training
Intel took a different path entirely. At Hot Chips 2026, Intel detailed Crescent Island, an AI accelerator built on its Xe3P architecture that skips high bandwidth memory in favor of up to 480GB of LPDDR5X, according to Tom’s Hardware’s reporting from the show. The chip packs 32 Xe3P cores, 256 XMX engines, 32MB of unified L2 cache, and runs over a PCIe Gen5 x16 interface at a 350-watt TDP for the air-cooled version, based on specs Intel confirmed and reported separately by TechSpot.
Intel will ship the standard configuration with 160GB of LPDDR5X, while ODM partners can build versions with up to 480GB. That’s a deliberate tradeoff. LPDDR5X is slower than HBM but far cheaper and easier to source right now, which matters enormously given the DRAM shortage described above. Intel’s bet is that long-context inference and agentic AI workloads, where a model needs to hold huge amounts of context in memory without necessarily needing HBM-class bandwidth, are underserved by GPUs built for training. Crescent Island is aimed squarely at that gap rather than at competing with Rubin or MI450 on raw training throughput.
What changed under the hood
Intel also redesigned the compute core itself. The new XMX units use a 16-deep systolic array, four times deeper than the previous generation’s four-deep design, and the Xe Vector Engine now supports FP8 and FP4 precision along with microscaling formats. Intel doubled the general register file to 1MB per Xe core as well. Individually these are incremental engineering choices, but together they signal Intel building specifically for the lower-precision, high-concurrency inference workloads that dominate production AI deployments today, rather than chasing the training benchmarks Nvidia and AMD compete on. Intel has promised Crescent Island for a second-half 2026 launch.
Head-to-head: how the three chips actually compare
| Vendor | Product | Memory | Target workload | Status |
|---|---|---|---|---|
| Nvidia | Vera Rubin (Vera CPU + Rubin GPU) | HBM-based, spec not fully disclosed | Training and inference, rack-scale | Production shipments started, AWS first |
| AMD | Instinct MI455X (MI450 series) in Helios racks | Not disclosed for this configuration | Training and inference, rack-scale | First gigawatt deploying H1 2027 |
| Intel | Crescent Island (Xe3P) | 160GB standard, up to 480GB LPDDR5X | AI inference, long-context and agentic | Promised for H2 2026 |
The comparison isn’t apples to apples, and that’s the point. Nvidia and AMD are both building rack-scale systems meant to handle training and inference at the largest hyperscalers, while Intel is deliberately targeting a narrower, cheaper inference niche where HBM’s bandwidth advantage matters less than raw memory capacity per dollar. Nvidia’s edge right now is that it’s already shipping Rubin in volume to a top-tier customer. AMD’s edge is a locked-in, multi-year customer commitment with equity alignment behind it. Intel’s edge is cost and memory capacity for a workload category that’s growing fast but still doesn’t need the most expensive silicon on the market.
Nvidia’s quarter versus the industry’s biggest commitments
| Metric | Figure | Source |
|---|---|---|
| Nvidia Q2 FY2027 revenue | $96.2 billion (+106% YoY, +18% QoQ) | Nvidia earnings release |
| Nvidia GAAP/non-GAAP gross margin | 75.0% | Nvidia earnings release |
| Nvidia Q3 FY2027 revenue outlook | ~$108 billion, plus or minus 2% | Nvidia earnings release |
| Nvidia FY2028 revenue growth guidance | ~70% | Nvidia earnings call, via CNBC |
| AWS GPU deployment expansion | +2 million GPUs | Nvidia announcement, Aug. 27, 2026 |
| AMD-Anthropic GPU commitment | Up to 2 gigawatts of Instinct MI450-series GPUs | AMD newsroom |
| AMD equity investment in Anthropic | Up to $5 billion | AMD newsroom |
| Gartner 2026 DRAM revenue forecast | +246.6% year over year | Gartner, via Data Center Knowledge |
Market impact: why the stock reaction wasn’t as simple as the revenue number
A 106% year-over-year revenue jump would normally be an unambiguous win, and in most quarters it has been for Nvidia. This time, the compressed gross margin drew more attention than usual, because it’s the first clear sign that memory costs are eating into Nvidia’s profitability rather than just its rivals’ pricing plans. Investors have spent two years treating Nvidia’s margin as close to untouchable. Seeing it hold flat at 75% while revenue doubled reads less like weakness and more like a company absorbing input costs to keep growing volume, but it’s still a data point worth watching next quarter.
The AWS expansion matters just as much as the earnings number for a different reason: it’s forward-looking proof that Rubin demand is real, not just guided. Public earnings guidance is a promise. A hyperscaler committing to 2 million additional GPUs on a new architecture, days after that architecture starts shipping, is closer to a receipt. That distinction is exactly why Nvidia’s stock tends to move more on infrastructure announcements like this one than on the headline revenue figure alone.
Historical context: from Hopper to Blackwell to Rubin
Nvidia’s Hopper architecture, built around the H100, powered the first wave of generative AI buildouts starting in 2023. Blackwell followed with B200 and GB200 systems, cementing the rack-scale approach that’s now standard across the industry. Rubin is the third architecture in that cycle, and it’s the first one to ship on an annual cadence rather than the roughly two-year gap between Hopper and Blackwell. That accelerated pace is a deliberate choice, and it’s also a risk. Every architecture Nvidia ships faster gives customers less time to fully depreciate the previous one, which works fine as long as demand keeps outrunning supply the way it has since 2023.
AMD’s Instinct line followed a slower but steadier path, moving from MI300 to MI350-series chips like the MI355X that Anthropic already runs, and now to the MI450 series. Intel, by contrast, spent most of the last three years without a credible answer in AI accelerators at all. Crescent Island is Intel’s first product built from the ground up around inference economics rather than trying to match Nvidia’s training performance, and that’s a meaningful strategic shift for a company that spent 2023 and 2024 mostly on the sidelines of this market.
The custom silicon wave adds a fourth front
Nvidia, AMD and Intel aren’t the only players shaping this market. OpenAI has confirmed its first custom AI chip, code-named Jalapeño and built with Broadcom, is a custom inference chip set to begin deployment in OpenAI’s own infrastructure by the end of 2026. We covered the Jalapeño benchmark claims and what they mean for Nvidia’s margins in a separate deep dive, and the short version is that OpenAI has said Nvidia still powers the vast majority of its inference fleet even with a custom chip on the way.
That pattern repeats across the industry. Google, Microsoft and Amazon have all built or expanded custom silicon programs of their own over the past two years, and Nvidia’s biggest customers are increasingly also its most credible long-term competitors, a dynamic we’ve tracked in our coverage of Nvidia’s customer-turned-rival problem. None of these custom chips are close to displacing Nvidia GPUs at scale today. What they do is give hyperscalers negotiating leverage and a hedge against any single vendor’s supply constraints, which matters enormously in a market where memory shortages are already rationing who gets what.
HBM4 and the memory supply chain behind all three chips
Rubin’s performance claims lean heavily on next-generation HBM4 memory, which reporting has tied to yield improvements that make higher-bandwidth memory more available for Nvidia’s newest platform, a topic we’ve covered separately. Intel’s decision to skip HBM entirely for Crescent Island looks less like a technical limitation and more like a hedge against exactly this kind of supply bottleneck. If HBM4 yields stay tight into 2027, Intel’s LPDDR5X approach could end up being the more available option for buyers who can’t get allocation from Nvidia or AMD, even if it can’t match their bandwidth ceiling.
Server vendors are already repositioning around this. HPE’s stock hit a 52-week high after leaning into Nvidia’s Vera CPU for its own server lineup, a move detailed in our earlier report, which shows how much of the server ecosystem is now built around Nvidia’s roadmap specifically rather than a generic x86-plus-GPU template. That concentration is exactly what AMD and Intel are trying to break with Helios and Crescent Island, respectively.
What this means for enterprise AI buyers
For a company deciding what to buy in late 2026, the practical takeaway is that there’s no longer a single default answer. Teams running large training jobs still have the strongest reason to wait for Rubin allocation or lock in Helios capacity through a cloud partner, since both platforms are built for that workload first. Teams running inference-heavy production workloads, particularly ones with long context windows or many concurrent agents, now have a real reason to evaluate Crescent Island once it ships, especially if memory allocation from Nvidia or AMD is constrained.
Pricing adds urgency to that decision. Consumer and workstation GPU prices have already climbed sharply this year on the same memory pressure hitting data center chips, and buyers who can shift some inference workloads to cheaper, LPDDR5X-based hardware get a real hedge against further price increases. The tradeoff is that Crescent Island won’t ship in volume until the second half of 2026 at the earliest, so anyone needing capacity right now still has to work within Nvidia’s or AMD’s supply constraints.
Predictions: where the AI hardware race goes from here
These are informed projections based on current trends, not confirmed roadmap details from any of the three companies.
- Memory supply, not GPU design, will stay the biggest constraint on AI hardware through 2027, and expect further price increases from all three vendors as DRAM and HBM allocation tightens further.
- Equity-linked supply deals like AMD’s Anthropic investment will become more common as chipmakers compete for guaranteed demand years ahead of shipment, rather than relying on purchase orders alone.
- Intel will lean harder into inference-specific benchmarks for Crescent Island rather than head-to-head training comparisons against Rubin or MI450, since that’s the workload category where its memory-capacity approach actually wins.
- Nvidia’s 70% fiscal 2028 growth guidance will face its first real stress test if the Rubin ramp slips even a single quarter, given how much of the current stock valuation already assumes that cadence holds.
- Custom silicon from hyperscalers, including OpenAI’s Jalapeño, will keep chipping away at the edges of Nvidia’s addressable market through 2027 without displacing its core training and inference business at scale.
Frequently asked questions
How much revenue did Nvidia report for its second fiscal quarter of 2027?
Nvidia reported $96.2 billion in revenue for the quarter ended July 26, 2026, up 106% from the same quarter a year earlier and up 18% from the prior quarter, according to Nvidia’s own earnings release.
What is Vera Rubin, and why did it ship to AWS first?
Vera Rubin is Nvidia’s newest data center platform, pairing a Vera CPU with a Rubin GPU. Nvidia confirmed production shipments began in August 2026, with Amazon Web Services as the first hyperscaler receiving systems at scale as part of an expanded deal covering 2 million additional GPUs.
How does AMD’s Anthropic deal compare to Nvidia’s AWS expansion?
AMD committed to deploying up to 2 gigawatts of Instinct MI450-series GPUs for Anthropic through its Helios rack-scale systems, alongside an equity investment of up to $5 billion. Nvidia’s AWS deal is a GPU volume commitment without a disclosed equity component. Both are multi-year infrastructure bets rather than one-time hardware orders.
What workloads is Intel’s Crescent Island built for?
Crescent Island targets AI inference, particularly long-context and agentic AI workloads that need large memory capacity more than the highest possible bandwidth. It uses up to 480GB of LPDDR5X memory instead of HBM, which Intel says keeps cost and availability better suited to inference-heavy deployments.
Why are Nvidia’s gross margins under pressure even though revenue doubled?
Nvidia’s GAAP and non-GAAP gross margins both held at 75.0% this quarter. The company has pointed to rising memory costs across its supply chain, driven by the broader DRAM and HBM shortage, as the main factor limiting further margin expansion despite record revenue.
Is the memory shortage affecting Nvidia, AMD and Intel equally?
All three are affected, but differently. Nvidia and AMD both rely on HBM for their flagship data center chips, putting them directly exposed to HBM supply constraints. Intel’s Crescent Island sidesteps HBM in favor of LPDDR5X, which is a deliberate hedge against that specific bottleneck, though it comes with lower memory bandwidth.
What is Nvidia’s revenue guidance for fiscal 2028?
Nvidia guided to roughly $108 billion in revenue for the third quarter of fiscal 2027, plus or minus 2%, and pointed to approximately 70% revenue growth for fiscal 2028 overall, according to commentary from the earnings call.
Will custom chips from OpenAI or other hyperscalers replace Nvidia GPUs?
Not in the near term. OpenAI has said Nvidia still powers the vast majority of its inference fleet even as it begins deploying its own Jalapeño chip in small volumes by the end of 2026. Custom silicon programs at Google, Microsoft, Amazon and OpenAI are growing, but none currently approach the scale of Nvidia’s data center business.
