Nvidia closed out August 2026 with the kind of week most chipmakers never get. On August 26, the company reported fiscal second-quarter revenue of $96.2 billion, up 106% from a year earlier, and paired that beat with a joint announcement alongside Amazon Web Services: AWS will deploy 2 million additional Nvidia GPUs across its global infrastructure through 2027 and 2028. The stock jumped nearly 9% the next trading day, adding roughly $440 billion in market value in a single session, according to CNBC. It was one of the largest one-day value gains any public company has posted this year.
The headline number is the GPU order, but the more interesting story sits underneath it: Nvidia is no longer just selling graphics chips. It is now shipping its first custom data center CPU, the Vera, at scale to hyperscalers including AWS, Oracle Cloud Infrastructure, and SpaceX AI. It has also introduced a new memory architecture, NVHBM, designed to squeeze more bandwidth out of a supply chain that is running short on high-bandwidth memory. And it is doing all of this while AWS, its biggest customer, keeps building a rival chip of its own. Here is what actually happened, what it means for the AI hardware race, and why AMD’s competing Helios platform makes this more than a one-company story.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
The Deal: 2 Million More Nvidia GPUs for AWS Through 2028
Nvidia and AWS announced on August 26, 2026 that the two companies plan to deploy 2 million additional Nvidia GPUs across AWS’s global infrastructure, spanning 2027 and 2028, according to Nvidia’s investor relations announcement. The rollout covers Nvidia’s newer Blackwell Ultra, Rubin, and Rubin Ultra chips, aimed squarely at workloads that did not exist as commercial products a few years ago: agentic AI systems that take multi-step actions on their own, enterprise automation pipelines, scientific computing, and physical AI for robotics.
The agreement goes beyond raw chip volume. AWS said it will offer Nvidia’s open Nemotron models through Amazon Bedrock and SageMaker, and will adopt Nvidia’s physical AI stack, Omniverse, Cosmos, Isaac, and Jetson, for its own warehouse robotics fleet. On the earnings call, Nvidia CEO Jensen Huang pointed to the 2 million GPU commitment and noted that AWS is also positioned to buy millions of Vera CPUs as that product ramps, tying the CPU business directly to the GPU deal rather than treating it as a side project.
Nvidia’s Fiscal Q2 2027 Earnings, By the Numbers
The quarter itself, covering the three months ended July 26, 2026, beat Wall Street on nearly every line. Total revenue of $96.2 billion topped the roughly $92.2 billion analysts had modeled, according to Reuters and Euronews. Data center revenue, the segment that matters most to the AI story, hit $89.0 billion, up 117% year over year and 18% from the prior quarter. Non-GAAP earnings per share came in at $2.22, ahead of the roughly $2.08 to $2.10 range analysts expected, while GAAP EPS reached $2.46. Gross margin held at 75.0% on both a GAAP and non-GAAP basis, a level few semiconductor companies of Nvidia’s size have sustained through a period of this much output growth.
| Metric | Fiscal Q2 2027 (ended July 26, 2026) | Change |
|---|---|---|
| Total revenue | $96.2 billion | +106% YoY, +18% QoQ |
| Data center revenue | $89.0 billion | +117% YoY, +18% QoQ |
| Non-GAAP EPS | $2.22 | Beat ~$2.08-$2.10 estimate |
| GAAP EPS | $2.46 | — |
| Gross margin (GAAP/non-GAAP) | 75.0% | Held flat quarter over quarter |
| Next-quarter revenue guidance | ~$108 billion | Above buyside estimates of ~$105B |
| Next fiscal year growth guidance | ~70% | vs. ~44% prior market consensus |
| Stock reaction (Aug 27, 2026) | +9% in one session | +$440 billion in market value |
A 70% Growth Forecast That Beat Wall Street’s 44% Consensus
The number that moved the stock wasn’t the quarter Nvidia had already booked. It was the one still ahead of it. CFO Colette Kress told investors Nvidia expects revenue to grow roughly 70% in the coming fiscal year, well above the roughly 44% growth analysts had penciled in before the call, based on reporting from Fortune and an earnings-note summary published by Gold River. Kress paired the bullish forecast with a warning: memory component shortages remain the binding constraint on how fast Nvidia can actually ship product, not customer demand.
That combination, demand outrunning supply rather than the other way around, is unusual for a company already generating data center revenue north of $89 billion in a single quarter. It also explains why Nvidia is racing to lock in memory supply. TrendForce reported on August 27 that Nvidia’s total supply commitments have climbed to roughly $279 billion as memory costs surge industrywide, a figure that dwarfs the annual revenue of most chipmakers on its own.
Wall Street’s Reaction: $440 Billion Added in a Single Session
Nvidia shares rose nearly 9% on Thursday, August 27, 2026, adding about $440 billion in market value in a single trading day, CNBC reported. For scale, that one-day gain alone is larger than the entire market capitalization of most Fortune 500 companies. The move reflected investors reading the 70% growth guidance, not just the quarter that had already closed, as confirmation that AI infrastructure spending has not plateaued.
Why Investors Bought the Guidance, Not Just the Quarter
Beating an already-high revenue estimate is common for Nvidia at this point. What is less common is a CFO raising forward guidance by roughly 26 percentage points above consensus while simultaneously flagging supply constraints. That combination told investors two things at once: the order book is deeper than the market assumed, and Nvidia still can’t build fast enough to satisfy it. Both read as bullish, which is part of why the stock move outpaced the earnings beat itself.
Inside Vera: Nvidia’s First Ground-Up Data Center CPU
The AWS deal isn’t only about GPUs. Nvidia is simultaneously ramping Vera, its first fully custom-designed data center processor, which the company has branded as the CPU built for AI agents rather than general-purpose computing. Vera packs 88 in-house Olympus cores and roughly 1.2 TB/s of memory bandwidth, and was formally unveiled at Nvidia’s GTC Taipei keynote on May 31, 2026, months before this week’s earnings call.
On the earnings call, Nvidia executives said Vera shipments are already underway to lead partners, explicitly naming Oracle Cloud Infrastructure and SpaceX AI, with AWS shipments beginning in the current quarter, according to a transcript published by Investing.com. Management’s stated goal is for Vera to be deployed by every major hyperscaler, neocloud provider, AI lab, and system OEM, positioning it as a direct challenge to the CPU businesses at Intel and AMD rather than a niche accessory to Nvidia’s GPU lineup.
NVHBM: Nvidia’s Answer to the Memory Bottleneck
Alongside the earnings and the AWS news, Nvidia published details on NVHBM, a new memory architecture that moves the HBM controller into the 3D memory stack itself rather than keeping it on the compute die. According to Nvidia’s developer blog and confirmed by TechPowerUp’s technical write-up, the change delivers up to 30% more memory bandwidth per stack compared with standard HBM4e, cuts HBM power consumption by up to 15%, and frees up to 25% more usable area on the compute die.
The timing isn’t a coincidence. Memory has become the binding constraint on the entire AI hardware industry in 2026, and squeezing more bandwidth and efficiency out of every HBM stack is one of the few levers Nvidia can pull without waiting on outside memory suppliers to add fab capacity. Notably, part of that technology is heading outside Nvidia’s own product line: Nvidia and Amazon’s chip design unit, Annapurna Labs, are working together to bring NVHBM to Amazon’s own Trainium chips.
The AWS Paradox: Buying Nvidia While Building Trainium
That last detail is the strangest part of this story. AWS is simultaneously Nvidia’s largest disclosed customer for this GPU order and one of its most direct long-term competitors, through its own in-house Trainium accelerator line. Rather than treating that as a conflict, the two companies appear to be formalizing a split: Nvidia GPUs and Vera CPUs for the most demanding frontier training and agentic workloads, with Trainium continuing to serve as AWS’s lower-cost, vertically integrated option for other customers, now enhanced with licensed Nvidia memory technology.
It is a pragmatic arrangement for both sides. AWS gets guaranteed access to the highest-performance silicon on the market at a moment when GPU supply is the tightest constraint in cloud computing, while continuing to control its own margins on Trainium-based instances. Nvidia, meanwhile, locks in a multi-year commitment from the world’s largest cloud provider even as that same provider keeps investing in an alternative to Nvidia’s own hardware. Readers comparing raw GPU rental costs across providers can see how wide that pricing gap already runs in our AWS vs Azure vs Google Cloud GPU pricing comparison.
AMD’s Countermove: Helios and the Instinct MI400 Family
Nvidia isn’t operating in a vacuum. On July 23, 2026, AMD unveiled Helios, its first rack-scale AI system, at the Advancing AI 2026 event in San Francisco, according to Moneycontrol. Helios bundles AMD’s new Instinct MI400-series GPUs with sixth-generation EPYC server processors (codenamed Venice), Ryzen AI Embedded X100 chips, and Kria AI robotics hardware into a single integrated portfolio, mirroring Nvidia’s own rack-scale strategy with Vera Rubin.
The flagship of that lineup, the Instinct MI455X, packs 432 GB of HBM4 memory and is being marketed by AMD as a direct challenger to Nvidia’s high-end platforms for frontier model training and agentic inference, according to a semiconductor weekly briefing from Distill Intelligence. AMD has also claimed that the related MI350P chip delivers up to 2.6 times more tokens per second per dollar than competing products in its class, an efficiency argument aimed at cost-sensitive inference buyers rather than the largest training clusters. For a fuller breakdown of how the two companies’ roadmaps line up, see our earlier look at Nvidia Rubin vs. AMD Helios vs. Microsoft Maia 300.
Nvidia vs AMD: Data Center AI Hardware Compared
| Category | Nvidia | AMD |
|---|---|---|
| 2026 flagship CPU | Vera: 88 Olympus cores, ~1.2 TB/s bandwidth | EPYC Venice (6th-gen EPYC) |
| 2026 flagship GPU line | Blackwell Ultra, Rubin, Rubin Ultra | Instinct MI400 series, led by MI455X |
| Flagship accelerator memory | NVHBM-enhanced HBM stacks | 432 GB HBM4 (MI455X) |
| Rack-scale platform | Vera Rubin (in production shipment) | Helios (unveiled July 23, 2026) |
| Efficiency claim | NVHBM: +30% bandwidth, -15% power vs. HBM4e | MI350P: up to 2.6x tokens/sec per dollar (AMD claim) |
| Disclosed hyperscaler commitment | AWS: 2 million GPUs, 2027-2028 | Not disclosed at comparable scale in current reporting |
| News trigger | Fiscal Q2 2027 earnings call, Aug 26, 2026 | Advancing AI 2026 event, Jul 23, 2026 |
How Nvidia Got Here: From Blackwell to a $96 Billion Quarter
It’s worth stepping back to see how fast this ramp has actually happened. A year earlier, in the equivalent quarter of fiscal 2026, Nvidia reported revenue of $46.7 billion, itself up 56% year over year at the time, per Nvidia’s SEC filings. This quarter’s $96.2 billion is more than double that figure. Data center revenue has followed the same trajectory, climbing from a business built almost entirely around Blackwell-generation GPUs a year ago to one now diversifying across GPUs, the Vera CPU line, networking, and now licensed memory technology heading to a competitor’s chip.
That diversification matters because it changes what kind of company Nvidia is being valued as. A pure GPU supplier lives and dies on one product cycle. A company selling GPUs, CPUs, rack-scale systems, networking, and licensed IP across the industry has more ways to keep growing even if any single product line slows down. The 70% growth guidance suggests investors, and Nvidia’s own finance team, believe that diversification is starting to show up in the numbers.
Market Impact: What This Means for Cloud Costs and AI Pricing
For developers and enterprises renting AI compute, the immediate effect of a deal this size is unlikely to be lower prices. Memory shortages, which Kress cited directly as a constraint on Nvidia’s own shipments, have already pushed up costs across the supply chain this year. Consumer GPU buyers have felt a version of the same pressure: Newegg pricing data tracked through late August showed Nvidia’s RTX 50-series desktop cards climbing as much as 39% since June, a trend we covered in detail in our report on Nvidia’s RTX price hikes. Enterprise GPU rental pricing has moved in a similar direction throughout the year as demand for training and inference capacity outstrips supply.
What the AWS deal does change is availability. A committed 2 million GPU pipeline through 2028 gives AWS customers more confidence that Blackwell Ultra, Rubin, and Rubin Ultra capacity will actually show up on schedule, rather than sitting on a waitlist behind hyperscaler-only allocations. For companies planning multi-year AI infrastructure budgets, predictable supply is often worth more than a marginally better price per GPU-hour, particularly for agentic AI and robotics workloads that are still early in their deployment curve.
The Memory Shortage Shadow Over the Boom
Every part of this week’s news traces back to the same underlying constraint: there isn’t enough high-bandwidth memory to go around. Nvidia’s $279 billion in supply commitments, reported by TrendForce, is as much a hedge against future HBM shortages as it is a demand signal. NVHBM exists specifically to extract more usable bandwidth from each memory stack Nvidia can actually secure, rather than waiting for memory suppliers to expand output. AMD faces the identical bottleneck with its 432 GB HBM4-equipped MI455X, which depends on the same limited pool of advanced memory suppliers Nvidia is drawing from.
That shared constraint is arguably a bigger long-term story than either company’s roadmap. Whichever chipmaker secures the most reliable memory supply, not necessarily whichever has the fastest individual chip, may end up controlling how quickly the rest of the AI industry can actually scale.
5 Predictions for Nvidia, AWS, and the AI Hardware Race Through 2028
- Vera CPU adoption broadens past the three named launch partners. With Oracle Cloud Infrastructure, SpaceX AI, and AWS already shipping, expect Nvidia to announce additional hyperscaler and neocloud commitments for Vera within the next two quarters, especially as Intel and AMD defend their own server CPU share.
- Memory suppliers, not GPU designers, become 2027’s most-watched earnings calls. With Nvidia’s $279 billion in supply commitments and AMD chasing the same HBM4 pool, expect memory makers to gain outsized pricing power and investor attention over the next several quarters.
- AWS keeps expanding Trainium even as it buys more Nvidia silicon. The 2 million GPU deal is unlikely to slow Amazon’s in-house chip roadmap. Expect AWS to keep positioning Trainium as its value tier while Nvidia GPUs and Vera CPUs anchor its premium, highest-performance offerings.
- AMD’s Helios ships its first customer racks in 2027. Given the July 2026 unveiling and Nvidia’s own Vera Rubin production ramp, AMD will face pressure to show real hyperscaler deployments of Helios and the MI455X within the next 12 months to keep pace with Nvidia’s disclosed AWS volume.
- NVHBM licensing extends beyond Trainium. If the Annapurna Labs partnership proves out performance gains on Trainium silicon, expect Nvidia to explore licensing NVHBM technology to additional custom-silicon programs at other hyperscalers, turning memory IP into a new, higher-margin revenue line.
Longer term, this week’s numbers reset the baseline for how the market judges good enough growth from Nvidia. A 70% forward guidance figure, delivered from an already-massive $96.2 billion quarterly base, makes it harder for any other AI infrastructure company, including AMD, to be judged on its own merits rather than against Nvidia’s trajectory. For a broader view of where custom silicon and merchant GPUs both fit in the AI chip landscape, see our AI chips 2026 hub.
Frequently Asked Questions
What did AWS and Nvidia actually announce on August 26, 2026?
The two companies said they plan to deploy 2 million additional Nvidia GPUs, including Blackwell Ultra, Rubin, and Rubin Ultra chips, across AWS’s global infrastructure during 2027 and 2028. The deal also covers Vera CPUs, Nvidia’s open Nemotron AI models coming to Amazon Bedrock and SageMaker, and AWS adopting Nvidia’s Omniverse, Cosmos, Isaac, and Jetson tools for its robotics operations.
How many Nvidia GPUs will AWS deploy, and by when?
2 million additional GPUs, on top of AWS’s existing Nvidia deployments, rolled out across 2027 and 2028 according to the joint Nvidia-AWS announcement.
What is Nvidia’s Vera CPU and who is buying it?
Vera is Nvidia’s first fully custom-designed data center CPU, built around 88 in-house Olympus cores and roughly 1.2 TB/s of memory bandwidth. It was unveiled at GTC Taipei on May 31, 2026, and is now shipping to lead partners Oracle Cloud Infrastructure and SpaceX AI, with AWS shipments beginning this quarter.
What is NVHBM and why does it matter?
NVHBM is Nvidia’s new memory architecture that moves the HBM controller into the 3D memory stack. Nvidia says it delivers up to 30% more bandwidth and 15% lower power than standard HBM4e, while freeing up to 25% more usable die area, at a time when high-bandwidth memory supply is the industry’s biggest bottleneck.
Why did Nvidia stock jump 9% after earnings?
Investors reacted less to the reported quarter, which beat estimates, and more to CFO Colette Kress’s forward guidance of roughly 70% revenue growth next fiscal year, well above the roughly 44% consensus analysts had modeled before the call.
How does AMD’s Helios and MI455X compare to Nvidia’s lineup?
AMD’s Helios rack-scale system, unveiled July 23, 2026, pairs the Instinct MI455X GPU, which carries 432 GB of HBM4 memory, with sixth-generation EPYC Venice CPUs. AMD has not disclosed a hyperscaler commitment at the same 2 million-unit scale that Nvidia confirmed with AWS, though AMD has made efficiency claims, including up to 2.6x more tokens per second per dollar for its MI350P chip.
Is AWS abandoning its own Trainium chips for Nvidia hardware?
No. AWS is expanding its Nvidia deployment while continuing to invest in Trainium, its own custom accelerator line. Nvidia is even working with AWS chip design unit Annapurna Labs to bring NVHBM memory technology to future Trainium chips, suggesting the two companies are settling into a coexistence rather than a replacement relationship.
Will this deal lower AI compute prices for developers?
Not directly, and possibly the opposite in the near term. Memory shortages, which Nvidia’s own CFO cited as a constraint on shipments, have already pushed both consumer GPU and enterprise GPU rental prices higher through 2026. The AWS deal mainly improves supply predictability for committed capacity rather than lowering per-GPU costs.
Related Coverage
- Nvidia RTX Spark Sinks Intel, AMD Stock 4.4% [2026]
- OpenAI's Mac Spree Rattles Nvidia's $5.5T Lead [2026]
- OpenAI Jalapeño Chip Beats Nvidia 1.9x on Power [2026]
- Nvidia RTX Prices Jump 39%, AI Chips Rise 15% [2026]
- Nvidia Reportedly Buys Hugging Face for $12.9B [2026]
