Nebius Swings to Profit as GPU Hunger Outruns Everything It Can Build

``` ---CONTENT_START--- The AI infrastructure market has a simple problem right now: too much money chasing too few chips. Nebius Group (NASDAQ: NBIS) is the latest company to prove that point with numbers that look less like a quarterly report and more like a growth-stage startup's pitch deck. The Amsterdam-headquartered AI cloud provider reported Q2 revenue of $582.3 million on Wednesday, up 454% from the same period last year and ahead of the $569.9 million consensus on TipRanks. The top line was strong enough to send NBIS shares up as much as 14% in pre-market trading, though the real story was in the margin profile. Adjusted EBITDA swung to a profit of $236.2 million, compared with a $21.0 million loss in Q2 2025. Adjusted net loss narrowed to $33.2 million—a 64% improvement year-over-year and dramatically better than the $190.7 million loss analysts had penciled in. The company ended the quarter with $8.04 billion in cash and cash equivalents, up from $3.68 billion at the end of 2025. But here's where the picture gets complicated. This is a company deploying capital at a pace that would make most CFOs physically ill. Capital expenditures hit $5.66 billion in the quarter alone, up from $510.6 million a year earlier. For the full year, Nebius expects to spend $20 billion to $25 billion on infrastructure. That's not a typo.

The Demand Side Is Doing Heavy Lifting

CEO Arkady Volozh, the former Yandex co-founder who stepped down from that company amid EU sanctions and was later delisted, has been characteristically direct about the current state of play. "Everything we do sells, and we are still in the very early stages," Volozh said during the Q1 earnings call. He repeated a variation of that theme in Q2 as demand continued to outstrip available capacity. The company has contracted power capacity of over 3.5 gigawatts, targeting at least 4 GW by the end of 2026, and remains on track to deliver 800 megawatts to 1 gigawatt of fully built and connected data center power by year-end. In practical terms, every GPU that comes online is already spoken for. Management has said that "usually four or more customers compete for every GPU" the company brings online, and that Q3, Q4 2025, and Q1 2026 capacity sold out in advance. Deferred revenue—a forward-looking indicator of locked-in customer commitments—stood at $5.98 billion, including $4.996 billion in non-current deferred revenue. That number is important because it reflects long-term contracts already signed, not aspirational pipeline. Overall backlog reportedly approaches $50 billion, with Microsoft and Meta anchoring the books. The Microsoft deal, signed in September 2025, is a five-year contract worth between $17.4 billion and $19.4 billion, running through 2031 and delivering dedicated GPU infrastructure from Nebius's new Vineland, New Jersey data center in phased deployments. Meta's commitment is even bigger: an initial $3 billion five-year deal in December 2025, expanded in March 2026 to up to $27 billion. That structure includes $12 billion in dedicated compute starting in 2027 plus $15 billion in additional on-demand capacity options. Meta is contractually obligated to purchase at least $12 billion of AI compute by 2027. Combined, Microsoft and Meta have committed more than $46 billion to Nebius through 2031.

The Nvidia Connection Runs Deep

Nvidia's $2 billion equity investment in March 2025 is worth a closer look, because it's structured in a way that reflects a highly strategic—some might say surgical—approach to locking in downstream compute demand. Nvidia holds 22,256,412 Class A shares, representing 9.3% of that class, according to a 13G filing disclosed on July 21, 2026. Only 1,190,476 of those shares were directly acquired. The remaining 21,065,936 were obtained through pre-funded warrants granted on March 11, 2026. Nvidia is contractually prohibited from exercising those warrants or selling the underlying shares before September 11, 2026. The 13G filing also carries a "passive investor" designation, confirming Nvidia has no board influence—this is a financial investment, not an operational partnership. One Nasdaq analysis described this as "a highly actuarial supplier-financing strategy designed to monopolize downstream compute demand." The lockup structure ensures maximum headline catalyst from the Nvidia endorsement while eliminating any immediate overhang on the stock. The chipmaker has also committed to helping Nebius build up to 5 GW of Nvidia-based computing infrastructure by 2030, and Nebius is deploying Nvidia's Blackwell Ultra platform across multiple new U.K. sites, representing a roughly $2.3 billion investment for 65 MW of compute capacity expected to come online from H2 2026.

The Bear Case Is Getting Louder

Michael Burry's short position on Nebius—disclosed August 6 at $211.77 per share—has drawn attention, though it's worth noting that Burry chose direct shorting over options because put premiums were too high with implied volatility exceeding 100%. His thesis centers on three points: AI infrastructure companies are depreciating chips too slowly, inflating earnings; data center execution carries massive risk given high costs and long lead times; and heavily indebted companies with long-term commitments could face severe distress if demand softens. Nebius reported $8.4 billion in debt as of March 2026. That number doubled during 2025. The company has since added a $775 million secured debt facility closed on July 17, 2026, with a maturity date of October 31, 2030, priced at SOFR plus 2.50% (roughly 6.8% at current rates). The facility was oversubscribed, led by MUFG with participation from ABN AMRO, Bank of America, Deutsche Bank, and HSBC. The execution risk Burry flagged isn't hypothetical. D.A. Davidson analyst Gil Luria visited Nebius's Vineland site and attended a city council hearing, then cut his price target from $250 to $175 on August 10, citing delays that could prevent the facility from completing construction this year. Vineland is critical to the Microsoft contract, and any slippage there threatens the $7 billion to $9 billion annualized run-rate revenue guidance by year-end. That's a meaningful risk. Insider selling has added to the bearish chorus. CEO Volozh sold 46,627 shares on July 1 at $235.45, netting approximately $11 million. Independent non-executive chair John Wilson Boynton IV sold 6,958 shares at $197 on July 15. Several other executives and directors also reduced positions during Q2, contributing to what one community analysis estimated at over $140 million in total insider sales during the quarter. Options traders had priced in a 13.2% move in either direction heading into the print. The premarket rally landed at the high end.

The CoreWeave Comparison Is Almost Too Easy

Nebius's closest public-market comp, CoreWeave, reported its own results earlier this week with Q1 revenue of approximately $2.1 billion, up 112% year-over-year. The two companies look similar from a revenue growth perspective but radically different on the balance sheet. CoreWeave carries approximately $51 billion in total liabilities against $2.2 billion in cash. Its Altman Z-Score sits at 0.3—deep in distress territory—compared with Nebius's 7.1, firmly in the safe zone. CoreWeave's debt-to-equity ratio is 739% versus Nebius's 132%. The company's current ratio is 0.3x versus Nebius's 8.3x. On the other hand, CoreWeave generated $5.98 billion in LTM operating cash flow versus Nebius's $2.84 billion. But its LTM free cash flow was negative $10.62 billion against Nebius's negative $3.15 billion. CoreWeave declined to provide explicit 2026 guidance; Nebius actually issued some ($3 billion to $3.4 billion group revenue). Piper Sandler named CoreWeave its top AI cloud pick with an Overweight rating and $151 price target, assigning Nebius a Neutral rating with a $224 target. Analysts there noted CoreWeave is "the better operating business today," but acknowledged Nebius's "lighter financial burden and focus on owned data centers could enable it to deliver better long-term returns." The market has made its own call on valuation. Nebius trades at a forward P/S of roughly 103x (Barchart data), against CoreWeave's 6.9x. But then again, going from $105 million to $582 million in revenue in a year is the kind of growth that commands a premium.

Wall Street Remains Divided

The analyst community can't agree on Nebius, and that's largely a function of how unprecedented this growth profile is. Goldman Sachs' Alexander Duval reiterated a Buy rating with a $286 price target ahead of earnings. Morgan Stanley initiated coverage with an Equal Weight rating and $126 target in January, then maintained Neutral with a $144 target after the Nebius Industry Summit. Heading into the print, TipRanks showed a Moderate Buy consensus with seven Buys and four Holds. The fundamental disagreement is about whether we're looking at a software company with hardware attached or a hardware company with software attached. Nebius's gross margin of roughly 74% in Q1 2026—the kind of number typically associated with SaaS platforms, not infrastructure providers—suggests the former. The company's Token Factory platform can cut inference costs for open models by up to 70%, with single-GPU inference gross margins reaching 81%. One reviewer on a community site gave Nebius AI Cloud a 4.93 out of 5 rating based on 222 reviews, noting the platform "rents NVIDIA H100 through GB200 clusters with managed Slurm, Kubernetes and an inference API." A Hacker News commenter observed that neoclouds like CoreWeave and Nebius are "using circular financing and massive debt to fund rapid AI infrastructure expansion, allowing hyperscalers to offload capital expenditures while securing early access to Nvidia GPUs."

The Path Forward Has No Playbook

There's no precedent for what Nebius is attempting. The company's capex plan represents a bet that demand will continue to outpace supply for years. Management says over half of the $7 billion to $9 billion ARR guidance is already contracted. The company is also exploring "light-asset partnership models" to expand capacity without fully funding every project, and had cash interest expenses of just $12 million annually against $236 million quarterly adjusted EBITDA. The financing structure is worth noting: a $4.3 billion convertible bond raise, Nvidia's $2 billion equity investment, $2.3 billion in quarterly operating cash flow from customer prepayments, and the $775 million secured facility. That's over $9 billion in 2026 financing. Management says 2027 customer commitments are already booked. The question isn't whether demand exists. It does. The question is whether a company growing this fast, spending this much, and trading at this multiple can execute flawlessly over the next three years. One delayed site in Vineland, one customer concentration risk (Microsoft and Meta represent the overwhelming majority of future revenue growth), one interest-rate shock, or one shift in hyperscaler procurement strategy could change the narrative quickly. Volozh's vision is to build the AI-native cloud platform that the hyperscalers can't build—where four layers (bare metal, multi-tenant cloud, Token Factory, agent layer) deliver cost advantages of 40% or more versus AWS, with an inference TCO that's 112% cheaper. That thesis is playing out in the numbers so far. Whether it holds at $25 billion of cumulative capex is another question entirely. The earnings webcast is set for today at 8:00 a.m. Eastern Time. Management says capacity expansion timelines, customer pipeline updates, and any changes to full-year guidance will be on the table. Given the stock's trajectory—still up over 129% year-to-date despite the July pullback—investors will be parsing every word about Vineland, about Meta's deployment schedule, and about whether those 2027 customer commitments convert into contracted revenue.