The pitch that built the GPU neocloud was simple and, for a while, true: renting an H100 from a company that does only GPUs is cheaper and faster than renting one buried inside a hyperscaler's 200-service menu. Skip the managed-service tax, get bare metal over InfiniBand, provision in hours instead of quarters. For 2023 and 2024, "cheap raw GPUs versus AWS" was the whole story.
That story broke in 2026, and the tell is embarrassingly direct: the best neocloud is the expensive one.
Price stopped being the axis#
CoreWeave is the reference neocloud — public on Nasdaq since March 28, 2025 at roughly a $23B valuation, the first cloud provider to deploy NVIDIA's GB300 NVL72, and the sole Platinum tier in SemiAnalysis's ClusterMAX 2.0 ranking of 84 providers. It is also, by repeated account, a premium vendor, not a discount one. When the highest-rated player in a market is the one you pay the most for, price is no longer the dimension buyers optimize.
The neocloud value proposition inverted while everyone was still quoting per-GPU-hour rates: the scarce input is no longer the silicon, it's the megawatts to run it.
What replaced price is availability — the unsexy physical question of who can get GB300 racks powered, cooled, and networked with non-blocking InfiniBand at gigawatt scale. That is a real-estate-and-electricity problem, and it reframes every provider as a balance sheet rather than a price list.
The numbers say "infrastructure fund," not "cloud vendor"#
Read the 2026 financials and the neoclouds look less like software companies and more like capital-intensive infrastructure plays:
- Nebius booked roughly $399M in Q1 2026 revenue, up ~684% year over year, against a contracted backlog north of $46B from Microsoft and Meta deals (per its SEC 6-K).
- Nscale raised a $2B round at a $14.6B valuation in March 2026 and contracted ~200,000 GB300 GPUs with Microsoft across Norway, Texas, and Portugal.
- Crusoe exited bitcoin mining entirely to become an "energy-first AI factory," raised a $1.375B Series E, and built the ~1.2GW Abilene campus anchoring OpenAI's Stargate.
- CoreWeave itself sits on roughly $22.4B of contracted commitments with OpenAI, assembled across three separate 2025 expansions.
The pattern is consistent: land an anchor tenant, use the contract to raise multi-billion-dollar debt, buy GPUs, repeat. Valuations track contracted backlog and power pipeline, not trailing revenue. Several analysts flag the circularity of it — NVIDIA-adjacent financing buying NVIDIA silicon to serve NVIDIA's largest customers — but as a buyer, the takeaway is narrower: the provider you pick is making a bet on power and capacity, and you are renting a slice of that bet.
The split that matters for agent builders#
Here is the part the "CoreWeave vs Lambda" framing misses entirely. The category is bifurcating into two businesses that barely compete.
On one side are the AI factories — CoreWeave, Crusoe, Nscale, Fluidstack — chasing a handful of mega-tenants with giant, contiguous, reserved InfiniBand clusters on the newest silicon. This is the right home for a large training run: you want cluster size, non-blocking topology, and reserved-capacity pricing, and you sign a multi-year commit to get them.
On the other side are the inference clouds — Together AI most explicitly, which just raised an $800M Series C at an ~$8.3B valuation on a serverless, per-token, agent-tuned stack. This is the right home for an agent fleet. Agent inference is bursty and latency-sensitive; it needs to autoscale to zero between spikes. Put that on a reserved bare-metal cluster and you pay for idle GPUs all night — the exact opposite of what agent economics can absorb.
That is why the naive head-to-head is the wrong question. Ask it differently:
- Training a model? You want a reserved InfiniBand cluster with the newest GPUs. Compare CoreWeave, Nebius, Crusoe, and Nscale on capacity, InfiniBand topology, and reserved price — availability of GB300 racks is the real constraint, not the hourly rate.
- Running an agent fleet? You want elastic, per-token, autoscaling inference. Compare Together AI and hyperscaler managed-inference tiers — and reserve raw neocloud capacity only for the steady, predictable baseline of your traffic.
The neocloud began as an arbitrage on hyperscaler pricing. It is ending as two separate industries: the landlords of powered silicon, and the toll-collectors on inference. The control plane keeps moving down into the hardware — and the money is moving with it. Pick the industry your workload actually belongs to before you compare a single price.



