Amazon's 2-Million-GPU Order Is a Supply-Chain Power Play, Not Just a Purchase

The Number Nobody’s Really Absorbing Yet

Amazon is ordering 2 million NVIDIA GPUs. That is not a typo. It is double the 1 million it originally committed to back in March at GTC.

The headline reads like another capex escalation story — which it partially is. But the real story is hiding in two details that mainstream coverage has largely ignored: the HBM customization clause and the timeline.

Both matter more than most readers realize.

The Order, Unpacked

AWS and NVIDIA announced Thursday that they will deploy 2 million additional GPUs across AWS’s global infrastructure through 2028. The chips will be Blackwell Ultra in 2027 and the upcoming Ruble and Ruble Ultra series in 2028. Alongside them come custom HBM (high-bandwidth memory) modules and NVLink Fusion interconnects — all co-engineered between the two companies.

This is not just buying more of the same chip from the same supplier. It is co-developing the memory and interconnect layers that determine whether those chips actually deliver performance at scale.

HBM is the bottleneck everyone talks about but few understand. Think of it this way: a modern GPU is only as fast as the memory feeding it. More compute cores without faster memory means idle silicon. The entire reason NVIDIA charges a premium over AMD or Intel in the data center is that its memory stack is tuned — custom-configured, custom-bundled, custom-timed — to extract performance that off-the-shelf components cannot match.

Amazon is now investing deeply into that custom stack alongside NVIDIA. That changes the competitive math.

Why Custom HBM Matters

HBM supply is not infinite. SK Hynix, Samsung, and Micron are the three manufacturers. Samsung has been chasing SK Hynix for the past two years and still trails. NVIDIA currently gives SK Hynix preferential allocations. Amazon’s custom HBM deal effectively locks AWS into a priority lane — one that Google and Microsoft cannot simply replicate.

Jensen Huang said on the conference call that NVIDIA expects fiscal 2028 revenue to grow roughly 70% year over year — far above the ~45% market consensus. That guidance implicitly acknowledges a demand curve that no single customer is driving alone, but Amazon’s order is the largest single commitment yet disclosed.

For context: AWS already committed to 1 million GPUs back in March. The fact that it doubled that number within months — while simultaneously signing up for custom HBM and next-generation interconnects — tells you something about the urgency inside Amazon’s machine.

The Internal Contradiction

Here is where it gets interesting. Amazon is also expanding its own AI silicon — Trainium processors are already in production, and the company has publicly stated it wants to reduce dependency on NVIDIA over time.

So why the massive GPU order? Because Trainium cannot yet handle the full workload. The current generation of AI training and inference demands NVIDIA’s software stack (CUDA) and its raw performance per watt. Amazon needs to scale now. It can diversify its silicon portfolio later.

This is a tension that will shape Amazon’s next several years: how much NVIDIA infrastructure do you buy while simultaneously trying to build a path away from NVIDIA? The answer, apparently, is: a lot.

What This Means for Google and Microsoft

Both are in the same position as Amazon — competing for the same GPUs, the same HBM allocations, the same datacenter space. Neither has publicly matched Amazon’s expanded commitment yet.

Google has its own TPU roadmap and has been relatively cautious about committing massive external GPU orders. Microsoft has been quieter about specifics but has been securing supply through partnerships with multiple vendors, including AMD and custom chip efforts via Xilinx (now part of AMD).

Amazon’s order is a first-mover advantage in the most constraining resource of all: HBM inventory.

If Blackwell Ultra and Ruble ships start arriving in 2027 with Amazon’s custom HBM, that gives AWS a performance edge that competitors must spend 12 to 18 months catching up on — assuming they can secure equivalent memory allocations.

The Timeline Is the Story

This is not a 2025 problem. It is a 2027–2028 problem.

The order places chips that won’t be widely available until 2027 in the Blackwell Ultra form factor, then Ruble and Ruble Ultra in 2028. That means AWS is committing capital today for infrastructure that will not fully deploy for at least two years.

Capital commitments this far out are unusual. They signal conviction that demand will keep accelerating — and that the window for securing supply is closing. Amazon is essentially saying: if I don’t lock this in now, I will be competing in 2028 for GPUs and memory I cannot get.

Jensen Huang’s Claim, Reconsidered

Huang called the current moment an “AI inflection point.” That phrase has been used before and will be used again. What is different this time is the commitment behind it. A 70% revenue growth forecast from NVIDIA, backed by Amazon’s 2 million GPU order, is not rhetoric — it is a financial bet on the direction of the industry.

If Huang is right, then the companies that secured supply now will have the infrastructure advantage for the next cycle. If he is wrong — if demand plateaus or fragmentation accelerates — then both NVIDIA and Amazon have overcommitted.

That risk is real. But the current trajectory does not suggest plateau. Demand for AI training and inference remains unsatisfied at scale, and the custom HBM deal Amazon has struck suggests it believes the next two years of growth are already baked in.

Who Wins, Who Loses

Amazon wins if this playbook holds: lock supply, co-develop custom memory, deploy at scale in 2027–2028, and use that infrastructure edge to capture enterprise and developer mindshare before competitors can react.

NVIDIA wins because it secures a guaranteed buyer for its highest-margin products across two full product cycles. It also strengthens its position in the HBM ecosystem by making Amazon a co-investor rather than a discount customer.

Google and Microsoft lose in the short term — not dramatically, but enough to matter. Every quarter Amazon gets ahead on infrastructure deployable capacity is a quarter it can undercut pricing, offer better SLAs, and attract workloads that competitors cannot yet serve.

The long-term outcome depends on whether Amazon’s own Trainium roadmap catches up fast enough to offset its continued NVIDIA dependency. For now, the bet is on NVIDIA. And the scale of that bet — 2 million GPUs, custom HBM, two-year timeline — is larger than most people realize.