NVIDIA and Amazon Web Services announced an expanded partnership on August 26, 2026 that pairs two separate but connected moves: AWS will deploy 2 million additional NVIDIA GPUs across its global infrastructure in 2027 and 2028, and NVIDIA’s open-weight Nemotron model family is now live on Amazon Bedrock and Amazon SageMaker. The announcement landed the same day NVIDIA reported fiscal second-quarter earnings, and the combination pushed NVIDIA shares up roughly 7% the following session, according to TradingEconomics and MarketBeat data. The deal is the clearest sign yet that NVIDIA’s open-model strategy and its GPU sales machine are now explicitly wired together.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What NVIDIA and AWS Actually Announced on August 26
The joint release, published on NVIDIA Newsroom, states that AWS plans to deploy an additional 2 million NVIDIA Blackwell Ultra, Rubin, and Rubin Ultra GPUs across AWS Global Infrastructure in 2027 and 2028. That figure comes on top of a separate commitment of more than 1 million GPUs announced at GTC 2026 for deployment beginning this year, according to reporting from TechCrunch. Put together, AWS’s publicly disclosed NVIDIA GPU commitment now tops 3 million units running from roughly 2026 through 2028.
TechCrunch’s reporting adds an important wrinkle: the earlier 1-million-GPU pledge from GTC 2026 was consumed faster than either company expected, with AWS burning through that allocation ahead of schedule as customer demand for NVIDIA-accelerated instances outpaced supply. That detail matters because it turns the new 2 million GPU figure from a marketing number into a capacity-planning response to a real shortage AWS was already experiencing.
Neither company disclosed a contract value. TechCrunch’s analysis notes that, based on per-unit list pricing for high-end AI accelerators running in the tens of thousands of dollars, the deal’s implied value likely runs into the tens of billions of dollars once spread across the 2027-2028 deployment window. That estimate is a third-party calculation, not an official figure from NVIDIA or AWS, and readers should treat it as directional rather than exact.
Nemotron Goes Enterprise: Inside the Bedrock and SageMaker Rollout
The infrastructure headline is only half the story. The same announcement confirms that NVIDIA’s Nemotron family of open models is now available on Amazon Bedrock as fully managed, serverless models, and on Amazon SageMaker for customers who want to deploy and fine-tune Nemotron on their own infrastructure. That two-track path — a plug-and-play managed endpoint on one side, a customizable base model on the other — mirrors how AWS has onboarded other third-party model families, but it’s the first time NVIDIA’s own open-weight lineup has received this level of native Bedrock integration.
This isn’t NVIDIA and AWS’s first pass at Nemotron distribution. Earlier in 2026, AWS’s Machine Learning Blog detailed how the Llama 3.3 Nemotron Super 49B V1 and Llama 3.1 Nemotron Nano 8B V1 models became available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart. The August 26 announcement extends that relationship to NVIDIA’s newest and fastest release, Nemotron 3.5 Lightning, and formalizes Nemotron as a standing model family on AWS rather than a series of one-off drops.
For enterprise buyers, the practical effect is that a company already running workloads on AWS can now call a Nemotron model through Bedrock’s API without provisioning GPUs itself, or pull the open weights into SageMaker to fine-tune a specialized version for tasks like code review, tool-calling agents, or customer billing questions — the exact use cases NVIDIA has marketed Nemotron 3.5 Lightning around since its release.
Nemotron 3.5 Lightning: The Model Behind the Push
NVIDIA released Nemotron 3.5 Lightning on August 11, 2026, alongside a routing library called NeMo Switchyard, according to SiliconANGLE’s coverage of the launch. The model is a 30-billion-parameter mixture-of-experts design with roughly 3 billion active parameters per token, built by distilling NVIDIA’s larger Nemotron 3 Ultra model. Its architecture interleaves Mamba-2 layers with MoE and select attention layers, a hybrid approach NVIDIA says delivers up to 4x the output speed of similarly sized dense models.
Crucially, Nemotron 3.5 Lightning is free for companies to download, use, and modify without paying NVIDIA or seeking permission, a point NVIDIA’s own technical blog and company blog both emphasize. SiliconANGLE’s reporting also cites a concrete customization example: CodeRabbit Inc. used NVIDIA’s standard auto-model recipe to train a specialized router agent for roughly $85 in about two hours, a figure NVIDIA is using to argue that Nemotron customization costs a fraction of training a model from scratch.
The economic logic is straightforward: NVIDIA gives away the model weights, then earns revenue when companies run those models — whether on their own NVIDIA GPUs, through NVIDIA NIM microservices, or now, through AWS’s Nemotron-on-Bedrock endpoints running on NVIDIA silicon. Every Nemotron download is, in that sense, a lead for GPU demand rather than a lost software sale.
How Nemotron 3.5 Lightning Compares to DeepSeek V4 and Qwen3.8
Nemotron 3.5 Lightning is not competing at the very top of the open-model leaderboard on raw reasoning benchmarks — it’s built for speed and cost at a mid-size footprint, not maximum capability. Independent benchmark tracking puts Nemotron 3.5 Lightning’s SWE-bench Verified score at roughly 51.6, with a GPQA Diamond score near 75.4 and an MMLU-Pro score around 81.9. Those are respectable mid-tier numbers, but they trail the largest frontier open models released this year by a wide margin.
DeepSeek’s V4-Pro Max, a much larger 1.6-trillion-parameter mixture-of-experts model with roughly 49 billion active parameters, scores around 80.6 on SWE-bench Verified and near 90 on GPQA Diamond. Alibaba’s Qwen3.8-Max, estimated at roughly 2.4 trillion total parameters, posts a GPQA Diamond score around 92.6 and an OSWorld-Verified score of 86.1. Even Alibaba’s smaller Qwen3.8-27B — closer in size to Nemotron 3.5 Lightning — outscores it on SWE-bench Pro and GPQA Diamond.
The takeaway isn’t that Nemotron is weak; it’s that NVIDIA built a different product for a different job. Nemotron 3.5 Lightning is optimized for throughput on narrow, repetitive agent tasks at a fraction of the compute cost of a trillion-parameter model, not for topping a general reasoning leaderboard.
| Model | Total / Active Parameters | SWE-bench Verified / Pro | GPQA Diamond | Positioning |
|---|---|---|---|---|
| NVIDIA Nemotron 3.5 Lightning | ~30B total / ~3B active | ~51.6 (Verified) | ~75.4 | Fast, cheap agent tasks; single-GPU capable |
| DeepSeek V4-Pro Max | 1.6T total / ~49B active | ~80.6 (Verified) | ~90.1 | Frontier reasoning and coding |
| DeepSeek V4-Flash | 284B total / ~13B active | ~79.0 (Verified) | Similar tier to Pro | Lighter frontier alternative |
| Qwen3.8-Max | ~2.4T MoE | 67.7 (Pro) | ~92.6 | Frontier general-purpose agent model |
| Qwen3.8-27B | 27B dense | 61.7 (Pro) | ~89.2 | Mid-size open alternative, outperforms Nemotron |
Benchmark figures are drawn from independent model-tracking summaries and NVIDIA’s own published documentation; treat cross-source benchmark comparisons as approximate since testing methodology varies by evaluator.
The 2 Million GPU Commitment, By the Numbers
The scale of the AWS order lands against a backdrop of already-stretched NVIDIA supply. NVIDIA’s Blackwell-generation chips have reportedly been sold out through mid-2026, with a backlog of several million units across the world’s largest cloud providers, based on industry supply-chain tracking. NVIDIA’s own quarterly disclosures reinforce the demand story: the company’s data center segment generated approximately $89 billion in revenue for its most recently reported fiscal quarter, up from roughly $75 billion the quarter before — an 18% sequential jump — and up sharply from about $39 billion in the same period a year earlier, according to earnings coverage cited by financial media following the August 26 report.
Consumer-facing GPU shipments tell a smaller but complementary story. Jon Peddie Research data reported by Windows Central put total discrete GPU shipments at roughly 11.8 million units in the first quarter of 2026, with NVIDIA accounting for around 10.7 million of those. The AWS deal sits on the data-center side of NVIDIA’s business, which now dwarfs consumer graphics as a share of total revenue.
Market Reaction: NVIDIA’s Stock and $5 Trillion Valuation
NVIDIA shares initially dipped roughly 1.6% on August 26 as the market digested a dense earnings report alongside the AWS news, then rallied hard the next session. By the close on August 27, NVIDIA traded in the $224-$226 range, up roughly 7% on the day, according to data separately reported by TradingEconomics and MarketBeat. That move pushed NVIDIA’s market capitalization to roughly $5.1 trillion, per CompaniesMarketCap tracking, keeping it the world’s most valuable public company.
MarketBeat’s extended-hours commentary on the earnings call singled out two catalysts for the rally: the AWS expansion itself, and the fact that early shipments of NVIDIA’s next-generation Vera Rubin platform have begun and are reportedly ramping faster than any prior NVIDIA product generation. Together, those two data points reassured investors that NVIDIA’s growth curve isn’t just holding — it’s extending into 2027 and 2028 on contracted, named-customer demand rather than speculative forecasts.
How the AWS-NVIDIA Deal Compares to Other 2026 AI Infrastructure Bets
The AWS-NVIDIA expansion is large, but it’s not happening in isolation. Every major hyperscaler and frontier AI lab has locked in multi-year infrastructure commitments over the past year, and comparing the raw numbers shows how the compute arms race has escalated since late 2025.
| Deal | Announced | Disclosed Value / Scale | Compute Detail |
|---|---|---|---|
| AWS – NVIDIA (2M GPU expansion) | Aug 26, 2026 | No official dollar figure; estimated “tens of billions” | 2M Blackwell Ultra/Rubin/Rubin Ultra GPUs, 2027-2028 |
| AWS – OpenAI (compute agreement) | Nov 2025, expanded 2026 | $38B initial, plus $100B expansion = $138B total | Hundreds of thousands of GB200/GB300 GPUs by end-2026 |
| Microsoft – OpenAI (Azure backlog) | Reported 2026 | ~$250B multi-year cloud commitment | Azure AI infrastructure capacity, unit count undisclosed |
| OpenAI – Oracle (Stargate, Abilene TX) | 2025-2026 buildout | ~$2.24B first wave estimate | 64,000 GB200 GPUs by end-2026, projected 450,000 over time |
| Anthropic – Google Cloud | 2026 | ~$200B five-year commitment | 10GW TPU/AI infrastructure, ~4M H100-class GPU equivalent |
Figures in the table are drawn from separate news reports and company disclosures across 2025 and 2026, several of which rely on analyst estimates rather than confirmed contract values; treat dollar totals as approximate. What the table makes clear is that the AWS-NVIDIA GPU count, while enormous in absolute terms, is one data point in a broader pattern: every hyperscaler is now locking in multi-year, multi-gigawatt AI compute commitments simultaneously, and NVIDIA silicon sits at the center of nearly all of them.
The Open-Weights, Closed-Cloud Strategy
What makes the Nemotron-on-Bedrock rollout notable isn’t the open license itself — Meta’s Llama, Mistral’s models, and Alibaba’s Qwen family are all open-weight to varying degrees. It’s that NVIDIA is pairing an open model with the tightest possible integration into a single hyperscaler’s proprietary cloud stack, at the exact moment that hyperscaler is buying millions of NVIDIA’s own chips.
A developer can still download Nemotron 3.5 Lightning’s weights and run it locally on a single desktop GPU — NVIDIA has marketed that lightweight deployment path explicitly. But the AWS integration creates a second, frictionless path where enterprises never touch the weights at all: they call a Bedrock endpoint, AWS bills them, and the underlying compute runs on NVIDIA hardware that AWS is now committed to buying in multi-million-unit batches. Open licensing removes the software toll; the hardware toll stays firmly in place.
Historical Context: From Nemotron Nano to a Planned Nemotron 4
Nemotron 3.5 Lightning didn’t appear out of nowhere. NVIDIA has been building out a structured Nemotron portfolio through 2026, including Nemotron 3 Nano, Nemotron 3 Super 120B, and Nemotron 3 Ultra — the last of which was released around May 21, 2026, at GTC Taipei, and was explicitly designed for long-running agent workloads with contributions from what NVIDIA calls the Nemotron Coalition. Nemotron 3.5 Lightning, released in August, is a distilled, faster derivative of that Ultra model rather than a from-scratch build.
Reporting has also surfaced NVIDIA’s next step: a Nemotron 4 family reportedly in development, with a flagship variant expected to reach at least 1 trillion parameters — roughly double the size of Nemotron 3 Ultra. Employees cited in that reporting say final training isn’t complete and no formal release date has been set, though a late-fall 2026 window has been floated. If that timeline holds, Nemotron 4 would be NVIDIA’s most direct attempt yet to match the scale of DeepSeek V4 and Qwen3.8 in the open-weight tier, rather than conceding the top of the leaderboard to Chinese labs.
The China Wildcard: Export Controls Still Shape the Math
NVIDIA’s ability to convert this kind of demand into shipped revenue remains partly constrained by U.S. export policy. In May 2026, the Commerce Department’s Bureau of Industry and Security clarified that license requirements for advanced AI chips apply to entities headquartered in China even when those entities operate outside the country. A separate rule that took effect in January 2026 allows case-by-case review for exporting H200-class chips to China, but caps such shipments at 50% of comparable U.S. export volume and requires third-party performance verification. Those restrictions don’t touch the AWS deal directly, since AWS’s GPU allocation is for U.S. and allied infrastructure, but they cap how much of NVIDIA’s total backlog can be filled by Chinese demand — keeping hyperscaler orders like AWS’s disproportionately important to NVIDIA’s growth math.
What Industry Sources Are Saying
NVIDIA and AWS’s own public statements frame the deal in structural terms rather than as a one-time order. In the companies’ joint announcement, NVIDIA described the expansion as building on “already-rapid customer adoption of NVIDIA-accelerated compute on AWS,” with the two companies planning to “deploy 2 million additional NVIDIA GPUs across AWS’s global infrastructure and deepen their work together across AI factories, CPUs, networking, open models, data processing and robotics, delivering co-engineered AI solutions that enable customers to accelerate AI development and deployment at unprecedented scale” — NVIDIA, in the official press release.
AWS’s portion of the same release specifies the timeline directly: “AWS plans to deploy an additional 2 million NVIDIA Blackwell Ultra, Rubin and Rubin Ultra GPUs in 2027-2028 across AWS Global Infrastructure, including AI factories” — AWS, in the joint announcement.
On the model-availability side, AWS confirmed the dual deployment path in the same statement: “NVIDIA’s Nemotron family of open models is available on Amazon Bedrock as fully managed, serverless models and on Amazon SageMaker for customers who want to deploy and fine-tune on their own infrastructure” — AWS, in the joint announcement. That framing was previewed months earlier when AWS’s Machine Learning Blog wrote: “Today, we are excited to announce that the Llama 3.3 Nemotron Super 49B V1 and Llama 3.1 Nemotron Nano 8B V1 are available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart” — AWS, on the AWS Machine Learning Blog.
Why This Matters for Enterprise Buyers
For CIOs and engineering leaders already on AWS, the practical calculus shifts in a few concrete ways. First, GPU capacity constraints that have made high-end NVIDIA instances hard to reserve on AWS should ease somewhat as the new allocation lands in 2027 and 2028 — though the prior 1-million-GPU tranche being consumed early suggests demand could absorb this new supply just as quickly. Second, teams evaluating open models for agentic workloads now have a lower-friction way to test Nemotron 3.5 Lightning without managing GPU infrastructure themselves, via Bedrock’s serverless pricing. Third, teams that need a customized, fine-tuned Nemotron variant can do so on SageMaker without waiting for NVIDIA to ship a purpose-built model, using the same auto-model recipe NVIDIA highlighted in the CodeRabbit case study.
The tradeoff is vendor concentration. A company routing agent workloads through Nemotron-on-Bedrock is now dependent on NVIDIA silicon, AWS uptime, and AWS pricing simultaneously — a tighter coupling than picking a cloud-neutral open model and hosting it independently.
Predictions: Where This Goes Next
- Expect Microsoft Azure and Google Cloud to announce comparable Nemotron integrations within the next two quarters, since NVIDIA has no incentive to keep an increasingly popular open model exclusive to one hyperscaler.
- Watch for a formal Nemotron 4 announcement in late fall 2026, likely timed to NVIDIA’s next major developer event, with a flagship model in the 1-trillion-parameter range designed to close the benchmark gap with DeepSeek V4 and Qwen3.8.
- The prior GPU tranche running out early suggests AWS may need to renegotiate or expand its NVIDIA allocation again before 2027 even arrives, especially if agentic-AI adoption accelerates faster than current forecasts.
- NVIDIA’s data center revenue growth rate will likely decelerate from the roughly 92% year-over-year pace reported this quarter, simply because the year-ago comparison base keeps rising — but absolute dollar growth should stay strong given the contracted GPU backlog already on the books through 2028.
- Export-control policy toward China will remain the single largest swing factor in NVIDIA’s total addressable market, and any further tightening or loosening in 2027 could move NVIDIA’s valuation more than any single hyperscaler deal.
Frequently Asked Questions
What did NVIDIA and AWS actually announce on August 26, 2026?
AWS committed to deploying 2 million additional NVIDIA GPUs (Blackwell Ultra, Rubin, and Rubin Ultra) across its global infrastructure in 2027 and 2028, and NVIDIA’s Nemotron open-model family became available on Amazon Bedrock and Amazon SageMaker, according to the companies’ joint press release.
How much is the AWS-NVIDIA GPU deal worth in dollars?
Neither company disclosed an official dollar value. TechCrunch’s analysis estimates the deal is worth tens of billions of dollars based on typical per-unit pricing for high-end AI accelerators, but that figure is a third-party calculation, not a confirmed contract value.
Is NVIDIA Nemotron 3.5 Lightning free to use?
Yes. NVIDIA has stated that Nemotron 3.5 Lightning is open-source and free for companies to download, use, and modify without paying NVIDIA or requesting permission, whether run locally on a single GPU or accessed through Amazon Bedrock.
How does Nemotron 3.5 Lightning compare to DeepSeek V4 or Qwen3.8?
Nemotron 3.5 Lightning is a smaller, faster model (roughly 30B total parameters) built for high-volume agent tasks, while DeepSeek V4-Pro Max and Qwen3.8-Max are much larger frontier models (1.6T and roughly 2.4T parameters respectively) that score higher on general reasoning benchmarks like SWE-bench Verified and GPQA Diamond.
Why did AWS need to order more NVIDIA GPUs so soon after its first 1-million-GPU commitment?
According to TechCrunch’s reporting, AWS’s earlier 1-million-GPU commitment from GTC 2026 was consumed faster than expected due to surging customer demand for NVIDIA-accelerated compute, prompting the larger follow-on order.
What is Nemotron 4, and when will it launch?
Nemotron 4 is a reportedly in-development NVIDIA model family with a flagship variant expected to exceed 1 trillion parameters, roughly double the size of Nemotron 3 Ultra. No official release date has been confirmed; some reporting has floated a late-fall 2026 timeframe.
How did NVIDIA’s stock react to the announcement?
NVIDIA shares dipped roughly 1.6% on August 26 as markets absorbed a dense earnings report, then rose approximately 7% the next trading session as investors digested the AWS expansion and strong data center revenue, pushing NVIDIA’s market capitalization to roughly $5.1 trillion.
Can I run Nemotron models without using AWS?
Yes. Because Nemotron models are open-weight, they can be downloaded and run independently of AWS, including on a single desktop or laptop GPU for Nemotron 3.5 Lightning. The AWS integration is an additional, managed deployment option rather than the only way to access the models.
