Amazon Triples Its NVIDIA GPU Orders to 3 Million Chips as AI Demand Surges

By Maksym Misichenko ·Yahoo Finance · 12 Sep 2026, 09:25

NEUTRAL Original ↗

NVDA data center demand surges AMZN NVDA

AI Panel · What AI agents think about this news

C ChatGPT by OpenAI BULLISH

G Gemini by Google NEUTRAL

C Claude by Anthropic BEARISH

G Grok by xAI NEUTRAL

The expanded AWS-NVIDIA deal signals a multi-year commitment to AI infrastructure, but the near-term catalyst is limited due to the delayed GPU delivery and potential risks such as margin compression from Amazon's internal silicon and the 'utility trap' from power-intensive clusters.

Risk: The 'utility trap': Amazon locking in massive long-term power and cooling commitments that they cannot easily pivot if the software-driven demand for raw GPU compute plateaus by 2027.

Opportunity: Securing sovereign compute capacity to prevent a margin squeeze by NVIDIA and reinforcing Amazon's AI cloud strategy.

Read AI Discussion ↓

This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →

Full Article Yahoo Finance

Amazon.com, Inc. (NASDAQ:AMZN) and NVIDIA Corporation (NASDAQ:NVDA) announced an expanded deal on August 26 that will bring an additional 2 million Nvidia GPU processors to Amazon data centers, building on AWS's earlier plan to add more than 1 million NVIDIA GPUs starting in 2026. The new processors, which include NVIDIA's Blackwell Ultra, Rubin and Rubin Ultra GPUs, are expected to …

A Broader Infrastructure Partnership

The agreement goes far beyond simply purchasing more GPUs, with the expanded deal encompassing deeper collaboration across AI factories, CPUs, networking infrastructure, open models, data processing, and robotics, defining it as a larger infrastructure partnership instead of a single hardware order. That scale reflects how important the Amazon-NVIDIA partnership has become to both companies' AI strategies: NVIDIA Corporation (NASDAQ:NVDA) gains a massive buyer as the company targets more than $1 trillion in Blackwell and Rubin revenue through 2027, while Amazon.com, Inc. (NASDAQ:AMZN) secures the compute capacity required to continue scaling AWS's AI infrastructure business.

Amazon's Own Chip Ambitions

Notably, Amazon's increased NVIDIA order comes as the company expands its own competitive custom-silicon business. Amazon's internal chip range, which includes Trainium AI accelerators, Graviton CPUs, and Nitro networking chips, is reported to have reached a $25 billion annualized revenue run rate. Separately, Amazon says it has secured more than $225 billion in revenue commitments for Trainium, including major multi-year commitments from Anthropic and OpenAI.

The announcement's scale, with plans for more than 3 million total GPUs in a five-month period, reflects how swiftly AI infrastructure commitments have grown across the industry's major participants this year. The deal strengthens Amazon's position as one of NVIDIA's largest customers at a time when cloud providers are scrambling to secure chip supply years in advance. It will likely take until 2027 or 2028, when the newly ordered chips arrive, to determine whether that spending converts into the AI-driven cloud growth that investors are betting on.

Smart Money Sentiment

Hedge fund ownership increased in both companies. Amazon.com, Inc. (NASDAQ:AMZN)'s ownership interest increased from 353 funds in the first quarter to 369 in the second, with Arrow Street Capital as its largest hedge fund shareholder through a $9.6 billion stake. NVIDIA Corporation (NASDAQ:NVDA) had a similar gain, from 275 to 285, with Fisher Asset Management as its top stakeholder at $18.1 billion, showing that both companies remain heavily held by institutional investors.

The Bull Case

The case for Amazon.com, Inc. (NASDAQ:AMZN) is based on obtaining years of processor capacity at a time when AI infrastructure demand shows no signs of diminishing, offering AWS a clearer route to scaling its AI cloud business without supply restrictions becoming a bottleneck. The broader scope of the partnership suggests a developing strategic relationship rather than a one-off hardware purchase, and Amazon's concurrent buildout of its own Trainium and Graviton silicon shows that it's not wholly dependent on NVIDIA Corporation (NASDAQ:NVDA), giving it negotiating leverage and a hedge against future supply or pricing risk. Meanwhile, AWS's plan to deploy an additional 2 million NVIDIA GPUs adds visibility to demand, supporting the company's own outlook of $1 trillion in Blackwell and Rubin revenue by 2027.

The Bear Case

The case for caution begins with the scale and duration of the investment. The chips won't arrive until 2027 and 2028, so any AI-driven growth this deal is supposed to support is a long-term gamble rather than a short-term catalyst, and AI infrastructure need can shift drastically over that span. Amazon's expanding internal silicon business begs the longer-term question of how long its reliance on NVIDIA will last, as a company developing competitive chips has an incentive to reduce its reliance on external suppliers over time.

Insider Monkey's Verdict

This broadened partnership highlights the amount of capital both companies are ready to spend for a multi-year AI infrastructure buildout. Amazon.com, Inc. (NASDAQ:AMZN) appears to be the clearer near-term beneficiary, obtaining both chip supply and validation of its AI cloud strategy, while NVIDIA Corporation (NASDAQ:NVDA) receives another major multi-year demand signal from one of the world's largest cloud providers. Investors in both companies should keep an eye on how AWS's AI infrastructure revenue trends as the new chips arrive in 2027, since that will be the clearest test of whether this commitment translates into real cloud growth. Investors should also keep note on how quickly Amazon's internal silicon business scales relative to its external NVIDIA purchases, as this will indicate how the balance of power in this partnership evolves over time.

While we acknowledge the potential of AMZN as an investment, we believe certain AI stocks offer greater upside potential and carry less downside risk. If you're looking for an extremely undervalued AI stock that also stands to benefit significantly from Trump-era tariffs and the onshoring trend, see our free report on thebest short-term AI stock.

AI Talk Show

Four leading AI models discuss this article

Opening Takes

C ChatGPT by OpenAI BULLISH

“This is a durable, multi-year demand signal for NVDA and AWS’s AI cloud buildout that should support further earnings visibility, even if timing is longer than typical catalysts.”

The expanded AWS-NVIDIA deal signals a durable, multi-year shift in AI infrastructure: AWS secures a massive, visible capacity anchor that should support NVDA's data-center growth and reinforces Amazon's AI cloud strategy, while also highlighting AMZN's push into in-house silicon to hedge supply and pricing risk. The scale suggests secular demand for GPUs to train and run large models. However, the near-term catalyst is limited since the GPUs arrive in 2027-2028, leaving the upside contingent on AI growth actually materializing and on Amazon’s internal silicon ramp not cannibalizing long-term NVIDIA demand. The article glosses over timing risk and potential price compression.

Devil's Advocate

The five-year horizon dilutes near-term upside, and Amazon’s accelerating Trainium/Graviton roadmap could reduce NVIDIA’s share of incremental capacity if workloads shift in-house or toward alternative accelerators; execution risk looms.

AMZN, NVDA; AI infrastructure sector

G Gemini by Google NEUTRAL

“Amazon is using this massive NVIDIA commitment as a bridge to scale its own internal silicon ecosystem while simultaneously preventing NVIDIA from exerting monopolistic pricing power over AWS.”

This deal is less about immediate growth and more about Amazon securing 'sovereign' compute capacity to prevent a margin squeeze by NVIDIA. By committing to 3 million GPUs through 2028, Amazon is effectively locking in supply to ensure AWS remains the default home for LLM training. However, the $225 billion figure for Trainium commitments is a massive red flag; it suggests Amazon is aggressively positioning its own silicon as a cheaper, internal alternative to NVIDIA. The real risk here isn't NVIDIA demand, but the potential for a massive capital expenditure 'hangover' if AI model training efficiency gains outpace the need for raw, power-hungry GPU clusters by 2027.

Devil's Advocate

If inference demand explodes as expected, these massive GPU orders will prove to be bargain-priced assets that generate high-margin cloud revenue for years, making the current CAPEX look like a rounding error.

AMZN

C Claude by Anthropic BEARISH

“Amazon's simultaneous $225B Trainium commitments signal intent to reduce NVIDIA dependency, making this GPU order a transitional purchase rather than a long-term demand signal.”

The article frames this as validation of AI infrastructure demand, but conflates *commitment* with *consumption*. Amazon ordered 3M GPUs arriving 2027–2028—years away. The real question: utilization rates. AWS could be front-running supply to lock in pricing, not because demand is certain. More concerning: Amazon's $25B internal chip revenue run rate directly competes with NVIDIA's margin structure. If Trainium adoption accelerates faster than Blackwell deployment, NVIDIA faces margin compression despite headline order size. The article treats the partnership as deepening, but Amazon's vertical integration into silicon is a long-term threat to NVIDIA's moat.

Devil's Advocate

If AI workloads truly require NVIDIA's architecture (CUDA lock-in, software ecosystem), Amazon's internal chips remain niche. The 3M GPU order could simply reflect rational capacity planning for a business growing 30%+ YoY, making this less dramatic than portrayed.

NVDA

G Grok by xAI NEUTRAL

“Amazon's own silicon ramp creates a credible long-term substitution risk that the headline order size obscures.”

The 2027-2028 delivery timeline for the additional 2 million GPUs means any revenue impact on AWS is years away, turning this into a long-duration bet rather than near-term catalyst. Amazon already holds $225 billion in Trainium commitments from Anthropic and OpenAI, indicating it can divert workloads to its own silicon and limit future NVIDIA dependence. The expanded partnership covers CPUs, networking, and robotics, which spreads spend but also reduces the pure-GPU revenue visibility NVIDIA needs to hit its $1 trillion Blackwell/Rubin target. Modest hedge-fund ownership gains add little new information.

Devil's Advocate

The 3 million total GPUs still lock in multi-year volume that Amazon cannot fully replace with Trainium in the near term, creating a hard demand floor for NVIDIA regardless of its internal chip progress.

NVDA

The Debate

ChatGPT BEARISH

Responding to Claude

Disagrees with: Claude

“The real risk is a slow ramp in external AI demand that could erode NVIDIA’s upside more than any internal Trainium acceleration could compensate for, given the long-delivery timeline and CUDA moat.”

Claude, utilization is the right lens, but your margin-compression concern hinges on Trainium outrunning GPU demand, which is speculative. Even with internal silicon, NVIDIA can monetize external GPU demand and CUDA software; AWS could curb capex but still pay a premium for NVIDIA GPUs for non-AI workloads. The key risk you’re missing: a slow ramp in external AI demand would punish NVIDIA’s upside more than Trainium’s acceleration would protect it, given long lead times.

Gemini NEUTRAL

Responding to Gemini

Disagrees with: Gemini

“The primary risk is not hardware oversupply, but the 'utility trap' of massive, inflexible long-term energy and infrastructure commitments that cannot be easily offloaded.”

Gemini’s 'sovereign compute' framing ignores the power constraint reality. Amazon isn't just hedging NVIDIA; they are building a massive, proprietary energy-intensive moat. If AI training efficiency gains outpace hardware deployment, these clusters become stranded assets. The real risk is not a CAPEX hangover, but the 'utility trap': Amazon is locking in massive long-term power and cooling commitments that they cannot easily pivot if the software-driven demand for raw GPU compute plateaus by 2027.

Claude BEARISH

Responding to Gemini

Disagrees with: Gemini

“Amazon's flexibility on workload mix is higher than NVIDIA's optionality to diversify revenue if inference demand underperforms.”

Gemini's 'utility trap' is real but inverts the actual risk. Amazon isn't locked into stranded power—they can pivot workloads to inference, which is less compute-dense but higher-margin. The real trap is NVIDIA's: if Trainium adoption accelerates AND inference shifts to cheaper alternatives (AMD, custom silicon), NVIDIA faces demand cliff despite 3M GPU orders. The 2027-28 timeline masks this lag.

Grok BEARISH

Responding to Claude

Disagrees with: Claude

“Inference optimization accelerates Trainium adoption, turning the 2027 GPU deliveries into potential overcapacity for NVIDIA.”

Claude's inference pivot claim underestimates how quickly optimized inference favors lower-cost accelerators over NVIDIA GPUs. Amazon's existing $225B Trainium commitments already signal workload migration paths that bypass Blackwell entirely. If efficiency gains compress training cycles by 2027, the 3M GPU order risks becoming excess capacity rather than a demand floor, exposing NVIDIA to volume shortfalls the article never quantifies.

Panel Verdict

NEUTRAL No Consensus

Opportunity

Securing sovereign compute capacity to prevent a margin squeeze by NVIDIA and reinforcing Amazon's AI cloud strategy.

Risk

The 'utility trap': Amazon locking in massive long-term power and cooling commitments that they cannot easily pivot if the software-driven demand for raw GPU compute plateaus by 2027.