Palantir and Nebius: a sovereign AI stack, not a merger of equals

Palantir and Nebius: a sovereign AI stack, not a merger of equals

The September 8 announcement that Palantir has named Nebius its preferred sovereign AI infrastructure partner is one of those deals that looks modest on a press-release page and large once you follow the logic. It is not an exclusive lock-in, it is not a co-sale of the two businesses, and it does not by itself crown anyone “the” market leader. What it does is pair Palantir’s software layer with a purpose-built GPU cloud inside Palantir’s own security perimeter — and that combination matters for how enterprises will buy AI over the next few years.

The deal is fresh. Palantir (PLTR) and Nebius Group (NBIS) said they will integrate Nebius compute and inference endpoints inside the Palantir enterprise perimeter after an integration period. Eligible commercial customers will then be able to run and continually adapt open models on Nebius infrastructure without pushing proprietary data and model weights out into a generic public cloud. The two firms will also try to bring capacity online faster, including modular data centres at sites that already have power.

Alex Karp put the thesis in one line: Nebius’s infrastructure lets customers run their own models under conditions they control; Palantir’s ontology and that infrastructure together “undergird the sovereignty our partners are demanding.” Arkady Volozh’s counterpart line was that organisations need both large-scale performance and control of data and models. That is the whole product.

Two different companies that suddenly need each other

Palantir is not a cloud company. It sells Foundry, Gotham, Apollo and, above all, AIP — the Artificial Intelligence Platform — plus the Ontology that maps an organisation’s objects, decisions and workflows so models can act on real operations rather than on a chat window. Its pitch in 2025–26 has been “sovereign AI”: AI that serves the institution instead of training a third-party frontier model on that institution’s data. US commercial revenue has been the proof point. In the second quarter of 2026, Palantir reported overall revenue of about $1.94 billion, up more than 90 per cent year on year, with US commercial revenue up 149 per cent to $764 million. Full-year guidance was lifted toward the $8.15 billion range, with US commercial expected above $3.42 billion. Those workloads have to run somewhere. Karp has spent the year arguing that token-based services quietly transfer enterprise IP. A preferred infrastructure partner that can sit inside the Palantir perimeter is the missing physical layer.

Nebius is the other half of that sentence, and its origin story is unusual even by AI-boom standards. The Nasdaq ticker NBIS is the renamed Dutch holding company that used to be Yandex N.V. After Russia’s invasion of Ukraine, trading was halted, Volozh was sanctioned and then unsanctioned, and in July 2024 the Russian consumer businesses were sold to a local consortium for roughly $5.4 billion. What remained — engineers, a Finnish data centre, cloud and AI infrastructure, plus stakes and subsidiaries such as Avride and TripleTen — became Nebius, an Amsterdam-headquartered “neocloud” selling GPU capacity as a service. Volozh returned as CEO. The company is not a 2023 garage startup; it inherited people who had already run large-scale search, maps, machine learning and cloud. That is why Palantir can claim Nebius was “built for AI from the ground up rather than adapted from general-purpose computing.”

The growth numbers explain why Palantir wanted that particular partner now. Nebius reported second-quarter 2026 revenue of $582.3 million, up 454 percent year on year, with the AI cloud segment growing even faster. Commentators cited an annualised run-rate around $3 billion, remaining performance obligations in the high tens of billions, customer commitments above $40 billion in some accounts, and a contracted-power target of 5 gigawatts. That is still smaller than CoreWeave’s backlog, but it is large enough to matter to Palantir’s commercial pipeline — and it is growing from a company that still has a European legal home and a narrative of control rather than hyperscaler lock-in.

Why this pairing, and why now?

The strategic logic is complementary, not overlapping. Palantir owns the authorisation, isolation, ontology and deployment layer. Customers already trust it with sensitive operational data. What they have lacked, in Palantir’s telling, is a sovereign option to train and serve their own open models on trusted metal without leaving that perimeter. Closed frontier models are expensive, general-purpose and, in Karp’s framing, extractive: you improve someone else’s model when you feed it your data. Open-weight models start weaker but can be looped on proprietary data until they beat the generic model on that domain. That only works if the compute is isolated, auditable and under the customer’s control. Putting Nebius endpoints inside Palantir’s perimeter is how you productise that claim for commercial accounts outside the US Army.

Nebius owns racks, power contracts, NVIDIA allocation and an AI-native software stack. Its problem has been distribution into the Fortune 500 and regulated industries that already standardised on someone else’s platform. Palantir’s forward-deployed engineers and commercial logo list are that distribution. A “preferred” badge from Palantir is also a credibility event for a firm that still has to live down the Yandex ancestry in some procurement rooms. The partnership is not described as exclusive. Palantir can still run on AWS, Azure, GCP, Oracle or on-prem. Nebius can still sell to everyone else. Preferred means first-call inside the sovereignty pitch, plus joint work on modular capacity where power already exists — a practical answer to the industry’s real bottleneck, which is no longer only GPUs but interconnects, transformers and megawatts.

The market’s first reaction was telling. Nebius shares jumped on the news; Palantir slipped. That is what you would expect if investors treat the announcement as a larger incremental demand signal for scarce AI compute than for a software vendor already priced as a winner. It does not mean the software side is unimportant. It means the scarce factor of production, for now, is still the factory.

What it does to everyone else in the stack

The relevant market is not “AI” as a blob. It is three layers that are colliding: hyperscale clouds, specialist neoclouds, and enterprise AI platforms. On the neocloud side, CoreWeave remains the scale leader by backlog and public profile, with contracted power measured in multiple gigawatts and a customer mix historically heavy on Microsoft and frontier labs. Lambda, Crusoe, Nscale and others compete on price, developer experience, energy strategy or geography. Nebius has been in that pack as a self-builder with European roots and a full-stack story. A Palantir preferred-partner designation does not make Nebius bigger than CoreWeave overnight. It does give Nebius a software channel the others don’t yet have at the same altitude. Expect copycat announcements: CoreWeave or Crusoe with a Databricks-, ServiceNow-, or Snowflake-shaped partner; European sovereign clouds courting the same Palantir customers; NVIDIA leaning harder on “sovereign AI” reference architectures that name more than one cloud. Validation of the neocloud category is real. Winner-take-all inside that category is not.

Hyperscalers should feel this more than they will admit in a blog post. AWS, Azure, Google Cloud and Oracle already host Palantir. They will keep doing so. The threat is narrative and procurement, not a sudden eviction. If a CIO can keep data, weights and fine-tunes inside a Palantir-controlled perimeter on Nebius metal, the default “just use our frontier model on our cloud” motion gets harder to defend in regulated industries, defence-adjacent commercial work, and any board that has started asking who actually owns the resulting model. Oracle has already played the sovereignty and dedicated-capacity card. Microsoft and Amazon will answer with more isolated regions, more confidential compute, and tighter AIP integrations. That is competition, not defeat.On the software side, Palantir’s peers — Databricks, Snowflake’s AI stack, C3.ai, Microsoft’s Fabric-plus-Copilot world, ServiceNow’s operational layer — are not made obsolete. They are being asked a sharper question: can you offer domain intelligence that the customer owns, or only rented intelligence that improves a platform vendor? Palantir’s bet is that ontology plus looped open models plus isolated compute is a different product from a lakehouse with a chatbot. The Nebius deal makes that bet cheaper to sell operationally. It does not prove the bet in production. Integration risk, capacity delivery and whether open models actually outperform closed ones after fine-tuning are still empirical questions.

Does this cement Palantir as the market leader?

Only if you define the market as Palantir already defines it. Palantir is the clear leader in a specific category: operational AI for complex institutions, with a government franchise that still funds the brand and a commercial engine that is now growing faster than the government book. No other listed software company has the same combination of classified-adjacent trust, forward-deployed implementation culture, and a working ontology that binds data to decisions. In that lane, this partnership strengthens the moat. It removes a practical objection — “where do we run our own models without leaking them?” — that hyperscalers were happy to leave unanswered. It does not make Palantir the leader of AI infrastructure. That title, such as it is, still sits with NVIDIA at the chip layer and with the hyperscalers plus CoreWeave at the capacity layer. It does not make Palantir the leader of foundation models. It does not freeze Databricks or Microsoft out of the enterprise. And “preferred partner” is a marketing and integration commitment, not a take-or-pay of Nebius’s entire 5 GW plan.

If Nebius misses build-out, or if Palantir’s commercial growth slows, the press release will look like a feature rather than a regime change. What it does cement is a direction of travel. The industry is splitting into two offers. One is rented intelligence: you send tokens to a closed model on a public cloud and accept that your data improves a shared brain. The other is owned intelligence: you take an open-weight model, loop it on your ontology-bound data, and keep the resulting advantage. Palantir has chosen the second offer and hired a factory to stand behind it. Nebius has chosen to be that factory for a software vendor that already has the accounts. Other players will have to pick a side or build a credible version of both. For investors watching both tickers, the clean reading is not “PLTR wins, everyone else loses.” It is that software and compute are bundling again, this time around sovereignty rather than around the original public cloud. Palantir looks more complete. Nebius looks more enterprise-grade. CoreWeave and the hyperscalers look like they need an answer that is more than another region and a confidentiality slide. That is a real shift in the market structure. It is not the end of the argument about who leads it.

By Anna Coulling – creator of volume price analysis

By Anna Coulling – creator of volume price analysis