If you want to understand where AI is actually going in 2026, stop watching model releases and start watching the infrastructure money. The numbers I’m seeing this quarter are unlike anything in the history of enterprise tech, and they’re being placed on a single set of bets: GPUs, data centers, and a handful of cloud operators willing to deploy them at scale.
The story this week is Nebius — a Dutch AI cloud provider most people outside the industry have never heard of. As of May 2026, Nebius has landed roughly $46 billion in AI cloud commitments, including a $27 billion deal with Meta and a $2 billion equity investment from Nvidia. To put that in context, Nebius didn’t even exist as an independent company three years ago. Today it’s one of the most strategically important AI infrastructure providers in the world.
Here’s what’s actually happening, and why it matters for your business.
The $700 billion number nobody is talking about
Combined 2026 AI capital expenditure across Amazon, Alphabet, Microsoft, and Meta is on track to land somewhere between $650 billion and $700 billion this year. That’s larger than the entire defense budget of the United States. It’s larger than the GDP of Switzerland. It’s a generational bet that the demand for AI compute will continue to outrun supply for at least the next 36 months.
Meta alone is projecting $115-135 billion in AI capex for 2026. That number was $30 billion just two years ago. The growth rate isn’t sustainable forever — but it doesn’t need to be. It just needs to last long enough for the AI companies to figure out how to monetize what they’re building.
Why Meta needed Nebius
Meta has its own data centers, its own custom AI chips, and a relationship with Nvidia that goes back further than most. So why is it spending $27 billion with a Dutch cloud company over the next five years?
The answer is timing. Meta’s internal AI infrastructure roadmap can’t move fast enough to keep pace with what Llama and the broader AI product portfolio need right now. Building data centers takes years. Securing power agreements takes longer. Nebius already has the locations, the power contracts, and — critically — direct allocations of Nvidia’s latest Vera Rubin chips, the next generation after Blackwell.
The Meta-Nebius deal breaks down to $12 billion in dedicated capacity across multiple Nebius locations, plus an additional $15 billion in committed compute capacity Meta can pull from over the term. The first deployment will be one of the first large-scale Vera Rubin clusters anywhere. That last part is the strategic kicker — being early on a chip generation matters more than the contract dollar figure.
And then Nvidia took an 8.3% equity stake in Nebius for $2 billion. That investment is the tell. Nvidia doesn’t take equity stakes lightly. When Nvidia buys 8% of an infrastructure provider, it’s not financial; it’s a signal about who Nvidia wants to be a long-term partner for the next chip generation. Other neoclouds — CoreWeave, Crusoe, Lambda — are watching this very carefully.
What this means for you (and the AI tools you use)
If you’re a startup founder or an enterprise IT leader, the Nebius story is a preview of three things that are about to land on your roadmap.
First, neocloud is becoming a real category. Until 2024, if you needed serious AI compute you went to AWS, Azure, or GCP. As of 2026, there are at least eight credible alternatives — Nebius, CoreWeave, Crusoe, Lambda, Foundry, Together, Fluidstack, and Ori — and they’re collectively eating the GPU rental market. Pricing on these platforms is typically 20-35% lower than the hyperscalers for equivalent workloads, with the trade-off being less mature managed services.
Second, GPU access is going to remain rationed for at least another year. The Vera Rubin generation is allocated almost entirely to hyperscalers, big neoclouds, and a handful of AI labs. If you’re a smaller AI company, your GPU pipeline for 2026 was probably set six months ago. New entrants are facing 12-18 month wait times for premium chips. This is going to push more AI workloads toward inference-optimized chips and specialized hardware.
Third, model prices are going to come down faster than people expect — but only for inference. Training costs are still going up because the frontier labs are buying compute as fast as it can be manufactured. But for inference, every additional gigawatt of capacity coming online drives prices down. Gemini 3.1 Flash-Lite at $0.25 per million input tokens isn’t an anomaly; it’s the new floor, and other model families will follow.
The risk nobody wants to discuss
I want to be clear-eyed here. The numbers being thrown at AI infrastructure assume that AI revenue will scale to justify them. We don’t yet know if that’s true. Anthropic, OpenAI, and the major model providers are all still operating at significant losses on the underlying compute economics. Enterprise AI revenue is real and growing, but it’s not yet growing fast enough to absorb a $700 billion annual capex line indefinitely.
If AI revenue growth slows in 2027, you’ll see one of two things happen. Either the hyperscalers absorb the overcapacity and turn it into cheaper general cloud — bullish for everyone using AI. Or they start rationalizing infrastructure investment, which would mean compute prices stop falling and could even reverse for a quarter or two. Watch the Q3 2026 earnings calls carefully — Meta and Microsoft will be the leading indicators.
What I’d do this quarter
Three moves if you’re building AI products today. First, lock in your inference pricing now if you’re at scale. Pricing volatility is going to increase as capacity comes online unevenly. Second, evaluate at least one neocloud — Nebius, CoreWeave, or Lambda — for any new AI workload you’re starting. The cost differential is real. Third, if you’re a serious enterprise buyer, ask your current cloud provider what their guaranteed Vera Rubin allocation looks like for late 2026 and 2027. Their answer will tell you a lot about which platform to standardize on.
The model wars get the headlines. The infrastructure wars decide who wins.
Stay on top of how AI infrastructure is reshaping the industry — see our latest AI News & Updates and the AI Tools built on top of it.
