Nebius: From GPU Lessor to AI Hyperscaler

Executive Summary

Nebius is the AI hardware buildout's most credible new entrant, and we think the market is still underwriting it as a GPU lessor rather than the hyperscaler it is becoming. NVIDIA's $2 billion strategic investment, custom high-density rack and cooling designs, and a pre-build capacity strategy already targeting hyperscaler and whale contracts collectively justify the 50% return since late 2024 (versus 25% for the S&P) and argue for further re-rating. The bifurcation we see most clearly: Nebius owns the physical layer of the AI cycle, while SpacetimeDB, now benchmarking at 303,920 TPS, is the software-layer analog (a unified realtime database and runtime) built for the agentic workloads that will run on top of that capacity. We lean long Nebius, and we treat SpacetimeDB as the most interesting software-layer optionality on the same secular tailwind.

The Business

Nebius sells AI compute capacity, but the underlying business is closer to a vertically integrated infrastructure builder than a reseller of someone else's GPUs.

Custom-Engineered AI Data Centers

Nebius designs its own racks, servers, cooling systems, and power distribution, optimized specifically for high-density GPU workloads. Per the company's own hardware disclosures, this yields meaningfully higher GPU density per square foot than traditional cloud deployments. That density matters because it directly compresses the two binding constraints on AI compute economics: real estate and power. As power-per-token becomes the dominant variable cost of inference, the operator with the lowest watts-per-FLOP captures durable margin advantage.

Pre-Build Capacity Strategy

Nebius is committing capital to data centers and GPU fleets ahead of signed contracts, a deliberately aggressive posture that trades balance-sheet risk for speed. In a market where hyperscaler and whale customers will sign with whoever can deliver capacity in quarters rather than years, ready inventory is the product. This is the same playbook CoreWeave used to win early ground, and Nebius is now running it with NVIDIA's capital backstop.

NVIDIA Full-Stack Partnership

NVIDIA's $2 billion equity commitment is not just capital. It is a technical alignment that gives Nebius preferred access to GPU allocations, reference architectures, and a co-engineered software stack. In a supply-constrained GPU market, that allocation pipeline is the single most valuable input a neocloud can hold.

Energy Efficiency as a Margin Lever

Nebius has explicitly oriented its hardware and software stack around lowering power per AI workload. As grid capacity becomes the binding constraint on hyperscaler expansion (already visible in Virginia, Dublin, and parts of Asia), the cloud with the best PUE and densest racks wins the marginal workload.

The Numbers

The clean market-data picture: NBIS trades near $135, with an approximate market cap of $16 billion on roughly 120 million shares outstanding. Since the late-2024 NVIDIA partnership baseline, the stock has returned roughly 50%, versus a 25% return for the S&P 500 over the same window, an outperformance of 25 percentage points. We estimate roughly $5 billion of incremental market cap is directly attributable to the NVIDIA partnership narrative and subsequent re-rating.

The structural numbers that matter:

• $2B NVIDIA strategic equity investment

• 303,920 ± 4,712 TPS SpacetimeDB TypeScript benchmark, a ~200% lift over the original 100,000 TPS number after V8 threading fixes

• TypeScript performance (303,920 TPS) now exceeds Rust (265,541 ± 940 TPS) on the same benchmark, an unusual inversion that signals the runtime, not the language, is doing the heavy lifting

The narrative that NBIS is "just" a GPU lessor is contradicted by the price action and the partnership terms. The market is starting to pay for hyperscaler optionality, not rental yield.

The Thesis

The bet is that AI infrastructure is bifurcating into two layers, and both are being re-rated simultaneously.

The hardware layer is consolidating around operators who can deliver dense, power-efficient GPU capacity at hyperscaler speed. Nebius checks every box: custom rack designs, pre-built capacity, NVIDIA capital and allocation, and a focused balance sheet that lets it move faster than incumbent clouds weighed down by legacy CPU footprints. The $2B NVIDIA commitment is the proof point that the hardware roadmap is credible. The 50% return since baseline is the proof point that the market is beginning to price it.

The software layer is where the next leg of the trade sits. Agentic workloads (LLM agents coordinating state, tools, and other agents) break the assumptions of traditional cloud backends. They need shared low-latency state, realtime sync across many participants, and a single runtime for logic and data. The standard answer (stitching Redis, Postgres, a queue, a sync service, and a serverless layer) is exactly the "glue" that LLM agents fail at most consistently. SpacetimeDB collapses that stack into a single unified database and compute runtime, and the 303,920 TPS benchmark says it can carry serious load.

What has to be true: AI inference demand keeps compounding at a pace that absorbs Nebius's pre-built capacity, NVIDIA continues to favor allocation to strategic partners, and agentic applications move from demos to production at scale. All three are directionally happening already.

Where We Differ

Consensus has Nebius slotted next to CoreWeave as a momentum-driven neocloud, with the implicit ceiling that neoclouds are GPU rental businesses with structurally lower multiples than the Big Three. We think that frame is wrong by one layer.

Nebius is not renting GPUs. It is engineering a custom AI hyperscaler stack, with NVIDIA's strategic capital and a pre-build strategy that mirrors how AWS scaled in 2009 (build ahead, sell into a demand wave). The market is comparing it to CoreWeave when the more accurate analog is early-stage hyperscaler economics with a Tier-1 chip partner. If that re-rating fully plays through, the multiple expansion is not 20% from here, it is multiples of the current price. The break above $130 suggests the market is just starting to see it.

The second-order divergence: most investors are not yet thinking about the software layer of this cycle. SpacetimeDB is private and early, but it is the cleanest read on what agentic backends actually look like in production. If you believe Nebius's thesis, you should also be tracking the software-layer winners that will run on top of that capacity.

The Open Questions

Can Nebius convert pre-built capacity into durable, high-margin direct demand?

The Bull View: NVIDIA's allocation pipeline, custom hardware density, and speed-to-deliver advantage mean Nebius will lock in multi-year contracts with both hyperscalers (as reseller capacity) and frontier AI labs (as direct customers). The pre-build risk pays for itself the moment the first whale contract signs.

The Bear View: Reseller deals with hyperscalers carry lower margins and limited pricing power. If direct enterprise demand is slower to materialize, Nebius ends up as a wholesale GPU supplier with thin margins and high leverage.

Our Read: The NVIDIA partnership is the tell. Strategic partners get strategic allocation, and strategic allocation gets sold direct to frontier labs, not wholesale to AWS. Nebius will land enough direct contracts in the next 12 to 18 months to validate hyperscaler-tier economics.

Does agentic infrastructure (and SpacetimeDB specifically) reach commercial scale?

The Bull View: LLM agents fundamentally need unified state, realtime sync, and one type system across client, server, and database. SpacetimeDB's 303,920 TPS benchmark and explicit positioning as the backend that "LLMs go much further with" is exactly the abstraction that wins when agents are doing the coding.

The Bear View: Enterprises adopt new backend architectures slowly. Postgres, Redis, and Kafka are entrenched, and SpacetimeDB has to displace muscle memory at the same time it earns trust on durability and operations.

Our Read: SpacetimeDB does not need to displace Postgres. It needs to be the default for the new category of realtime agentic and multi-user applications being built from scratch. That is a meaningfully smaller hurdle, and the LLM-native positioning is well-targeted.

Is the NBIS valuation already pricing the hyperscaler outcome?

The Bull View: At ~$16B market cap, NBIS is a fraction of CoreWeave's valuation and an even smaller fraction of what a real hyperscaler trades at. The re-rating has barely started.

The Bear View: A 50% return in roughly a year, on a name that does not yet have hyperscaler-scale revenue, embeds aggressive expectations. Any deceleration in AI capex commentary hits this name first.

Our Read: The valuation embeds growth, but it does not yet embed hyperscaler classification. The next leg comes from category re-rating, not earnings beats.

Risks

Pre-Build Capacity Underutilization

Committing capex ahead of contracts is the single largest balance-sheet risk. If AI training demand cools or hyperscalers internalize more compute (Anthropic on Trainium, OpenAI on custom silicon), Nebius could be holding dense GPU capacity at depressed utilization. The magnitude is meaningful: data center capex is multi-year and largely fixed.

Pricing Compression in Neocloud

CoreWeave, Lambda, Crusoe, and a long tail of regional neoclouds are all chasing the same workloads. As capacity comes online industry-wide through 2026, GPU-hour pricing likely compresses. Nebius's density advantage helps margin, but it does not immunize them from headline price pressure.

Valuation Sensitivity to AI Narrative

NBIS has outperformed the S&P by 25 points since the NVIDIA deal. That outperformance is narrative-driven and will reverse hard on any meaningful AI capex disappointment. A single guidance cut from a hyperscaler customer hits this name harder than the index.

SpacetimeDB Adoption Friction

The software-layer thesis depends on enterprises and developers actually adopting a new unified runtime. Benchmark performance is necessary but not sufficient. Adoption headwinds, competition from cloud-native incumbents, and the slow pace of backend re-platforming all argue for a multi-year, not multi-quarter, path.

What Would Break This

• NBIS GPU utilization falls below ~70% on a reported basis through 2026, signaling pre-built capacity is not clearing

• NVIDIA reduces strategic allocation to Nebius or signs a competing partnership of similar scale with a rival neocloud

• A hyperscaler customer publicly cuts or delays a Nebius reseller commitment, breaking the multi-year demand narrative

• NBIS fails to hold the $130 level on a sustained basis through year-end 2026, signaling the hyperscaler re-rating has stalled

• SpacetimeDB benchmark numbers are independently challenged or fail to translate into named production deployments by mid-2026

• AI inference pricing per token compresses faster than density gains, eroding the energy-efficiency margin lever

The Takeaway

We land long Nebius as the cleanest pure-play on the AI hardware buildout, and we view SpacetimeDB as the most credible software-layer analog on the same cycle. The single most important reason: NVIDIA does not commit $2 billion of strategic capital to a GPU lessor, and the market is still in the early innings of re-rating NBIS as the hyperscaler it is becoming. Watch the next hyperscaler-tier direct contract announcement: that is the data point that closes the gap between "neocloud momentum trade" and "emerging hyperscaler" in consensus framing.