Locked In: What $1 Trillion in AI Compute Capital Means for Your Infrastructure Decisions

Commissioned, Curated and Published by Russ. Researched and written with AI.

What’s New This Week

At GTC San Jose (March 16-21, 2026), Jensen Huang announced that Nvidia now sees at least $1 trillion in high-confidence demand and purchase orders across Blackwell and Vera Rubin through 2027 – up from the $500 billion figure he cited at GTC 2025. The DGX Station GB300 opened for pre-orders during the same week: a deskside system with 784GB of coherent memory, 20 petaflops at FP4, and a price point aimed at engineering teams that want to stop renting inference capacity. Vera Rubin, the Blackwell successor, was confirmed in full production with product shipments planned for the second half of 2026.

Changelog

The number that came out of GTC this week is $1 trillion. That’s the figure Jensen Huang put on expected purchase orders for Nvidia’s Blackwell and Vera Rubin chip platforms through 2027. It was $500 billion last year at the same conference, for a shorter window. Now it’s doubled, and the horizon has extended.

Most of the coverage treated this as a valuation story. It isn’t, or at least that’s not the interesting part. The interesting part is what it means when capital at this scale is already committed and being manufactured – and what that implies for the infrastructure decisions engineering teams are making right now.

The Irreversibility Point

Purchase orders at this scale don’t get cancelled. Fabs are running. Supply chains are locked. The chips will exist. The question isn’t whether Nvidia will ship $1 trillion worth of compute between now and the end of 2027 – it’s what happens to the market once that compute is out in the world.

This matters because the usual counterargument to AI infrastructure investment is “what if the hype cools?” That argument becomes less persuasive when the physical infrastructure to support the next capability wave already exists and is already being paid for. The models will get better because the compute to train them is already in production. You can be sceptical about AI sentiment without being sceptical about that.

Cloud Compute Costs Are Not Going Down

The intuitive read on “more supply” is “cheaper prices.” That’s probably wrong in the near term, and here’s why: the $1 trillion figure reflects demand that is tracking ahead of supply, not behind it. Huang’s framing at GTC was that computing demand has increased by a million times and that the orders they’re seeing reflect that scale of appetite. Hyperscalers are committing to enormous volumes – Amazon alone reportedly has roughly one million chips committed through the end of 2027.

When demand scales faster than supply, cloud AI compute costs don’t fall. They stay high, or they fall slowly while the hyperscalers work through their backlogs. If your cost model for an AI workload assumes that GPU-hours will get materially cheaper in 2026, you should revisit that assumption. The arithmetic doesn’t support it.

The On-Premise Inference Case Gets Stronger

The DGX Station GB300 is the most concrete data point for the other side of this argument.

It’s a deskside box. It has a 72-core Grace CPU and a Blackwell Ultra GPU connected via NVLink at 900 GB/s, giving you 784GB of coherent memory across the whole system. The GPU side is 252GB of HBM3e at 7.1 TB/s bandwidth; the CPU side is 496GB of LPDDR5X at 396 GB/s. Peak compute is 20 petaflops at FP4 with sparsity – which is a marketing number, but even with the usual caveats it represents a step change in what you can put under a desk.

What that 784GB figure actually means: you can run a 1-trillion-parameter model locally without quantising it into uselessness. That has not been possible on a single deskside system before. The memory ceiling that has made large model inference a cloud-only problem is lifting.

The economics shift when you compare a one-time hardware cost against ongoing cloud inference spend at scale. The crossover point depends on your workload volume and latency requirements – but the crossover now exists for more teams than it did two years ago. If you haven’t modelled it for your specific inference workload, it’s worth doing.

For teams that want inference performance without a full deskside commitment, Nvidia also announced the RTX PRO 4500 Blackwell Server Edition – a single-slot 165W card that Nvidia claims delivers 100x faster inference compared to CPU-only setups. Server-rack inference that doesn’t require specialised rack infrastructure.

Vera Rubin Is Not a Rumour Anymore

Vera Rubin (the Blackwell successor) was confirmed at GTC as being in full production across seven chips in the platform, with product shipments scheduled for H2 2026. Nvidia’s performance claims are significant – the company points to meaningful inference improvement over Blackwell, with token cost reduction figures cited in the announcements leading up to GTC.

The production confirmation matters because it removes one category of uncertainty: this isn’t a roadmap slide. The manufacturing is underway. Teams evaluating infrastructure timelines for 2027 should treat Vera Rubin as a known quantity, not a maybe.

Vendor Lock-In Risk Has a Different Profile Now

A vendor that has $1 trillion in committed purchase orders is not going anywhere. That’s a different risk profile than building on Nvidia’s ecosystem three years ago, when the company’s dominance was real but the moat was less obvious.

The flip side: CUDA, NIM microservices, and Nvidia’s Foundry platform represent significant switching costs if you build deep into them. The argument for platform commitment is now stronger – the vendor is more durable – and the argument for portability is also stronger – the lock-in is more permanent. Neither cancels the other out. They mean the decision deserves more deliberate analysis than it did before.

If your team is currently building on CUDA-native tooling, the question isn’t whether to stop – CUDA is the only practical option for most workloads at scale. The question is whether your inference serving layer is portable enough that a future hardware shift, if one ever comes, doesn’t require a full rewrite.

What To Actually Do With This

Training is still hyperscaler territory. If you’re training large models, nothing in this week’s announcements changes your infrastructure calculus – you’re renting time on someone else’s $1 trillion bet.

Inference is where the decision space is opening up.

The concrete things worth doing now:

Model your inference costs at volume. Take your current or projected inference workload and run the numbers against dedicated hardware – DGX Station or server-grade Blackwell cards – versus cloud GPU instances. The calculation may or may not favour on-premise; that depends on your utilisation patterns and tolerance for capital expenditure. But the calculation is now worth doing in a way it wasn’t two years ago.

Revisit that model when Vera Rubin pricing becomes clearer, likely late 2026. Generation-on-generation improvement in inference efficiency changes the crossover point. The hardware available in H2 2026 will be materially better per dollar than what’s available now.

Be honest about your portability assumptions. If your inference stack is deeply coupled to a specific provider’s tooling, that’s a risk worth acknowledging explicitly – not necessarily acting on, but acknowledging.

The $1 trillion is already committed. The question is whether your infrastructure roadmap is treating that as a fixed variable.