What was this call, exactly?
A fireside chat, not an earnings call. Nebius presented at the Goldman Sachs Communacopia + Technology Conference in San Francisco, and the session is available as a 33-minute audio replay. No slides, no new financial statements, and no guidance update. What it offered instead was the clearest management statement yet of how Nebius intends to sell, price, and finance capacity through 2028.
| Event | Goldman Sachs Communacopia + Technology Conference, September 8 to 11, 2026 |
| Nebius session | September 8, 2026, 3:45 p.m. Pacific Time, 33 minutes, audio replay |
| Speakers | Arkady Volozh, Founder and CEO; Marc Boroditsky, Chief Revenue Officer |
| Moderator | Alex Duval, Goldman Sachs, Head of Europe Technology Hardware and Semiconductors Research |
| Replay | Recorded webcast, also linked from the Nebius investor hub |
One housekeeping note before the substance: the moderator opened by calling the session off the record, yet Nebius links the replay from its own investor hub and a full transcript circulates publicly. We are summarizing that public replay, in our own words.
How far out does the demand visibility now run?
Further than management was willing to say a quarter ago. Volozh's framing, in sequence:
The window moved. Where the company previously talked about roughly 18 months of demand visibility, he now puts it at 24 months, maybe more. The driver is simple: AI is producing measurable value in coding, security, and office work, and that value flows down the chain to the model builders and then to the compute providers.
Orders are arriving for 2028. Customers are now asking for capacity for the first and second quarters of 2028, in blocks of tens of thousands of GPUs. Nebius is still not pre-selling most of it. The policy from the second-quarter call stands: build the capacity first, sell it later, closer to delivery, at prices the market sets then.
He used the word unlimited, with a caveat attached. Demand, in his telling, exceeds anything the physical world can build on every visible horizon, while conceding that nothing is actually unlimited. Read it as a statement about the order book in front of him, not a law of nature.
What does the Palantir sovereign partnership actually do?
The partnership was announced the morning of the conference, September 8, and it was the session's news item. Palantir named Nebius its preferred sovereign AI infrastructure partner. After an integration period, Nebius compute and inference endpoints will run inside the Palantir enterprise perimeter.
Volozh explained the underlying idea in plain terms. An enterprise sitting on proprietary data has two choices. Send that data to a commercial model and improve someone else's model, potentially including a competitor's. Or run an open-weight model under its own control, train it repeatedly on its own data, and end up with a model that, inside that company's narrow domain, can beat a general commercial model. He pointed to Shopify as a worked example of the second path, where an open model tuned on company data outperformed the frontier commercial models it replaced.
Palantir supplies the software loop that does the training and governance. It previously ran that loop on hyperscaler infrastructure, which meant client data left the client's control. Nebius supplies a controlled stack, from the data center and racks up through the cloud to Token Factory, its inference platform running open models. Palantir controls the stack end to end for its customers.
| Layer | Who provides it | What the customer gets |
|---|---|---|
| Sovereign software and governance | Palantir (AIP, Ontology, Foundry, Apollo) | Authorization, isolation, and the data loop |
| Models | Open-weight models, tuned on customer data | Domain intelligence the customer owns |
| Inference platform | Nebius Token Factory | Day-zero support for frontier open models |
| Compute and data centers | Nebius AI Cloud | GPU capacity inside the Palantir perimeter |
Boroditsky's contribution was the commercial read: Palantir's credibility and installed base in the enterprise give Nebius a channel into customers it could not reach quickly on its own, and a route to the customer diversification the company keeps promising. Both men were careful, and investors should be too: no dollar value, no term, no capacity figure, and no named customer accompanied the announcement. It is a distribution agreement with strategic logic, not backlog.
Why are customers still buying older GPUs?
Because their workloads fit them. Boroditsky pushed back on the assumption that demand for older generations is desperation buying. Most workloads running on older chips, retrieval systems, text prediction, image generation, were built and tuned for those chips. The total cost is known, the reliability is proven, and the businesses on them keep growing. Nebius keeps a waiting list of new customers for older-generation capacity as it frees up.
On top of that base, demand for Blackwell is running at a multiple of the chips Nebius can supply, and customer interest is already projecting forward to Vera Rubin. The generational ladder, in other words, is not cannibalizing itself. Each rung has its own queue.
What are the three deal buckets, and what do they pay?
Boroditsky recapped the contract architecture that the second-quarter call introduced. The Communacopia version adds the strategic role of each bucket; the pricing references are from the second-quarter disclosures, which the conference comments were consistent with.
| Bucket | Term | Role in the model | Reference economics |
|---|---|---|---|
| Short-term | 3 to 6 months | Reserved slice of capacity for time-boxed training and post-training sprints, where weeks of delay cost the customer more than the premium | $40M to $50M per MW (Q2) |
| Medium-term (core) | 1 to 3 years | The lion's share. AI natives, digital natives like Shopify, and early enterprise adopters. Where Nebius wants the business to live | $20M to $25M per MW (Q2) |
| Long-term | 5 years and longer | Hyperscaler-scale contracts taken deliberately for their financing value: contracted cash flows that support asset-backed borrowing to fund the core build | Funding tool first, margin second |
The four landmark core deals from the second quarter, averaging more than $1B each with Reflection, Cohere, a scaled US neolab, and a large US quant trading firm, were won in the proof of concept, not on price. One customer's technical team reportedly called it the best evaluation experience they had had with any supplier. All four are already in discussions for additional clusters, including Vera Rubin capacity.
What did the capacity auction prove?
That Nebius did not know its own clearing price, and the market's was higher. Management first tried raising prices manually. Demand did not blink. So the company ran an experimental auction for Blackwell capacity in August, and the clearing price came in 15 to 20% above the highest price Nebius had ever charged, after all its manual increases.
Two details make this more than a pricing anecdote. First, the winners were pleased: in a shortage, an auction is the one mechanism that guarantees access if you win, and the winning bidder said so publicly. Second, Nebius intends to extend the experiment: more capacity into auctions, and auction-style mechanisms, including preemption, for its large self-serve customer base. Price discovery is becoming a product feature, not a negotiation.
The postscript arrived after the conference: on October 13, Nebius disclosed GPU price increases of more than 30% against first-quarter levels. The auction was not a one-off print.
Should investors worry that Meta and xAI become competitors?
Volozh's answer was the session's most dismissive, and it came with arithmetic. A year ago, the world's AI capacity additions ran at roughly 5 to 10 gigawatts a year. Next year and in 2028, he expects 15 to 20 gigawatts. Against that, xAI's talked-about 6 to 8 gigawatts is a 20 to 30% increment, welcome but not market-breaking. Meta has not actually launched a cloud offering; xAI has sold capacity, but at the bare-metal level, to buyers like Anthropic, Google, and Azure, who then put their own cloud on top.
His punchline: that bare-metal layer is exactly what Nebius wraps. If Meta or xAI has spare capacity, Nebius would happily take it, repackage it, and resell it to end users. New supply, in this framing, is inventory, not competition. The honest caveat is that this holds while end demand absorbs everything built. It is a shortage argument, and it expires when the shortage does.
How does Nebius grow beyond the customers it already has?
Three routes, all described on the call:
Up the customer ladder. From AI natives (media names like Higgsfield and HeyGen were cited) to larger AI natives signing bigger, longer deals, to digital natives embedding AI in their products, and now to platform providers that serve the enterprise, several of which are treating Nebius as their first non-hyperscaler supplier. Enterprise, management believes, is ultimately the majority of the market. The focused verticals: physical AI, healthcare and life sciences, media and entertainment, retail, and financial services.
Up the stack. Customers who arrive for training compute are being sold inference on the same platform, then agentic building blocks: Tavily, the agentic search company Nebius acquired, is described as a cornerstone, with blueprints today and fuller agentic services to come. This is the stated path from GPU hours to hyperscaler-style value capture and margin.
Sideways, through partners. That is the asset-light model, next.
What is the asset-light model, in plain terms?
Power and data center companies are trying to move up the stack toward the AI margins, and discovering they lack the three things Nebius has: the know-how to design and build GPU data centers, the racks and software stack, and the customers. Many of them do have land, power, and access to cheap financing.
The partnership splits the work accordingly. The partner finances and owns the physical asset. Nebius supplies the design, its racks, its full-stack platform, and its go-to-market demand. Management described a long list of candidate partners, each one a separate large project, with capacity realistically arriving in 2027 and later. Volozh called it a non-linear way to grow alongside the company's own build pace. For investors, the appeal is revenue and margin without Nebius's balance sheet carrying every gigawatt, the same logic DalalBytes flagged when the model debuted in the second quarter.
On power itself, the closing answer was pure scarcity logic: grid-connected plots, behind-the-meter generation, and partners like Bloom Energy for on-site generation are all in use, and the company will take any power it can get, because everything it builds is already sold.
What did the call not tell us?
This is the section that matters for the stock, because the market has moved violently since September 8 while the fundamental script has not:
No numbers. No guidance change, no backlog update, no Palantir economics, no financing announcement. The last hard scoreboard remains the second quarter:
| Q2 2026 scoreboard | Value |
|---|---|
| Revenue | $582.3M, up 454% year over year |
| AI Cloud revenue (98% of group) | $574.9M, up 514% |
| Exit ARR (end of June) | $3.0B |
| Group adjusted EBITDA / margin | $236.2M / 41% |
| Operating cash flow (prepayment-driven) | $2.3B in the quarter |
| Cash | $8.0B |
| Q2 capex | $5.7B |
| FY26 guidance (reaffirmed) | Revenue $3.0B to $3.4B; exit ARR $7B to $9B; capex $20B to $25B |
| Contracted power target, year-end | 5 GW |
No answer on funding mix. The Communacopia session never revisited how the 2026 capex bill gets paid if prepayments slow or credit tightens. Our Nebius investment report scores the company 70 out of 100, Positive with a funding watch, and the watch is unchanged by this call. In this week's AI power trade ranking, Nebius placed first of seven on growth and survivability combined, with the explicit condition that prepayments hold above roughly half of new-deal capex.
The price context has flipped. NBIS traded near $249 on the Palantir announcement day. It closed at $165.47 on October 16, 2026, down 12.18% that session in a sector-wide rout on rate and credit worries, at roughly a $42B market value. Management's September story and October's price are both real. They are answers to different questions.
What is the DalalBytes take?
Communacopia was a strategy call, and as strategy it coheres. Sell in three buckets and let the long contracts fund the core. Auction scarce capacity instead of guessing its price. Take the enterprise through Palantir's door rather than building a sales force door by door. Let partners' balance sheets carry some of the gigawatts. Each piece reinforces the others, and each was described with unusual operational specificity for a conference chat.
But a fireside chat cannot retire the only question the market is asking, which is financial. Nebius is attempting one of the largest private infrastructure builds in history on a $3B revenue base, funded substantially by its own customers' prepayments and by capital markets that have turned hostile this month. Volozh is right that demand is not the constraint. He has been right about that for five straight quarters. The constraint is the cost and availability of the money, and on that subject September's most confident call said nothing at all. The third-quarter report, whenever it lands, has to.
Related reading: the Nebius five-pillar analysis, and The Inference Stack, on where Token Factory sits in the post-training build-out.
Frequently asked questions
What did Nebius announce at the Goldman Sachs Communacopia conference?
No new financial guidance and no new contract values. In a 33-minute fireside chat on September 8, 2026, CEO Arkady Volozh and Chief Revenue Officer Marc Boroditsky extended the demand visibility window from about 18 months to 24 months or more, said orders are now arriving for the first half of 2028, explained the Palantir sovereign AI partnership announced the same day, and said the first Blackwell capacity auction cleared at prices 15 to 20% above anything Nebius had charged before.
What is the Palantir and Nebius sovereign AI partnership?
Palantir named Nebius its preferred sovereign AI infrastructure partner on September 8, 2026. After an integration period, Nebius compute and inference endpoints will sit inside the Palantir enterprise perimeter, so Palantir customers can train open-weight models on their own data, on infrastructure they control, without feeding that data to a commercial model provider. The companies disclosed no dollar value, contract length, or capacity figure, and Nebius presented it as an enterprise channel rather than a booked contract.
Why does the funding question still hang over Nebius?
Because the build is far larger than the revenue base today. Nebius guides to $20B to $25B of 2026 capital expenditure against $3.0B to $3.4B of revenue. The model works while customers keep prepaying: about 70% of second-quarter deals carried upfront payments covering 50 to 60% of the associated capex, and management expects more than $9B of prepayments in 2026. If prepayments or capital markets tighten, the same demand story has to be funded with more debt or more shares. NBIS closed at $165.47 on October 16, 2026, down 12.18% that day, which is the market pricing the funding risk, not the demand.
Disclosures
Research and opinion, not investment advice. The author holds long positions in Nebius (NBIS) and Palantir (PLTR). Positions may change without notice. Figures are drawn from the public conference replay, company releases, and filings believed reliable but not guaranteed; verify before acting. DalalBytes research on Nebius: investment report and five-pillar analysis.
© 2026 DalalBytes Research ยท For educational purposes only. Not investment advice.