Nebius Is Raising the Price of Its AI Compute on Oct. 1. Here's What That Says About the Shortage.

Nebius will raise Oct. 1 on-demand prices for four Nvidia GPU models by 17%-21%, taking B300 pricing to $9.50 per hour from $7.85; H100 and B300 rates are roughly 53% and 56% above May levels. Management says it could sell all 2027 capacity today under its newest contract terms, while annual contract value per megawatt has risen from about $12 million to more than $20 million, with short-term deals reaching $40 million-$50 million per MW. AI-cloud adjusted EBITDA margin expanded to about 50% in Q2, but the company plans $20 billion-$25 billion of capex this year and has funded expansion partly through $5.75 billion of convertible notes and roughly $2.8 billion in equity. The pricing increase signals acute AI-compute scarcity, though its direct near-term revenue effect is limited because much capacity is under pre-priced reserved contracts and the stock already trades above 20x midpoint 2026 sales guidance.

Analysis

The economically important signal is not the posted-rate change but the implied scarcity rent on unsold incremental megawatts. NBIS has already contracted a substantial portion of forward output, limiting near-term revenue repricing; therefore, valuation upside requires continued high-priced bookings, on-time power/data-center delivery, and sustained utilization rather than simply stronger spot pricing. This makes the equity increasingly a duration-and-execution asset: a small delay in energized capacity or a shift in contract mix can have an outsized effect on equity value given the capital intensity.

NVDA benefits at the margin if elevated rental economics preserve cloud-provider returns on new accelerator deployments, reducing the risk of a near-term order digestion cycle. PLTR's infrastructure partnership is strategically useful for sovereign workloads, but it is not yet evidence of material incremental PLTR revenue; the more relevant read-through is that customers may accept bundled software/infrastructure architectures to secure scarce capacity, favoring platform vendors with deployment credibility.

The key second-order risk is synchronized supply. Hyperscalers, CoreWeave-type GPU clouds, and sovereign projects are all funding capacity against the same shortage signal; spot and short-duration pricing are usually the first margins to normalize once power interconnects and GPU deliveries catch up. NBIS also faces refinancing and dilution sensitivity: capital returns can look exceptional at peak pricing while residual asset values and customer concentration become problematic in a down-cycle. Over 1-3 months, track new contracted MW and financing terms; over 6-18 months, the decisive variable is whether realized revenue per MW remains above the level required to earn through depreciation, interest, and equity dilution.

Contrarian view: scarcity is likely real, but public-market pricing may be capitalizing it as durable. The bull case needs a widening moat in power access, customer acquisition, or sovereign compliance—not merely access to Nvidia hardware, which competitors can also procure. A rate rollback, lower auction clears, a material increase in unutilized capacity, or a downward revision to EBITDA margin/payout assumptions would falsify the durable-scarcity thesis quickly.

Key Decisions for Investors

• Avoid adding NBIS on the rate announcement alone; treat it as a booking-quality watch item. Reassess after the next earnings release only if contracted MW, revenue per MW, and utilization support forward pricing without incremental equity issuance.
• For existing NBIS exposure, reduce or hedge into strength using 3-6 month put spreads if available; the asymmetric risk is multiple compression from capacity timing, financing, or spot-price normalization rather than a modest earnings miss.
• Maintain NVDA as the cleaner long exposure to sustained GPU-cloud scarcity over 6-12 months: demand durability improves accelerator order visibility while NVDA bears materially less project-finance and utilization risk than NBIS. Revisit if major cloud customers signal capex digestion or Blackwell lead times normalize sharply.
• Do not underwrite PLTR upside from the infrastructure partnership until management identifies contract value, deployment milestones, or sovereign-customer conversion. Use any partnership-driven rally as an alert to compare incremental PLTR bookings against valuation, not as a standalone catalyst.
• Monitor GPU-cloud comparables and disclosed auction/contract pricing monthly; a sequential decline in short-term revenue per MW or evidence that new capacity is landing ahead of contracted demand is a trigger to turn tactically negative on NBIS over the following 1-3 months.