For those holding Nvidia or considering AI infrastructure investments, a key question has emerged: AI money isn't shrinking—it's moving elsewhere. The rental price of Nvidia's B200 chip has fallen from a high of $6.11/hour on May 30 to $4.22/hour as of last weekend, a drop of about 30% in three weeks. At the same time, a rare divergence has appeared in the semiconductor sector: the SMH semiconductor ETF rose 15% over the past month, Micron and SanDisk each surged nearly 60%, while Nvidia fell about 3%.
B200 Rental Price Drops 30% in Three Weeks, Putting Pressure on the 'Compute Scarcity' Narrative
The Nvidia B200 is the core compute chip for hyperscale data centers, and its rental price is considered a barometer of AI infrastructure supply and demand. Ornn data shows the B200 hourly rental price has steadily declined from its May 30 high of $6.11 to $4.22. AIMultiple's monthly price index compiled from 63 cloud providers shows a median B200 quote of $6.11/hour, but neocloud vendors have pushed the floor down to $3.44/hour. GetDeploying, which tracks 26 B200 cloud providers, reports even more extreme figures: an average price of $4.99/hour and a lowest quote of just $2.25/hour (for a three-year reserved contract).
Three factors are driving the price decline: improved yield at TSMC's 4NP process reduces B200 production costs; SK Hynix and Micron's HBM3e supply has loosened significantly in Q2 2026; and more neocloud providers have secured B200 inventory, with RunPod, Lambda, Nebius, Spheron, and others now offering spot availability, intensifying competition and lowering overall prices. Pressure will increase further in the second half of the year. As Nvidia's next-generation Blackwell Ultra B300 enters the spot pool, some B200 capacity will shift from on-demand to spot pricing. B300 spot prices have already been seen as low as $2.45/hour, cheaper than the lowest B200 listing. Spheron and Thunder Compute predict the B200 on-demand price could stabilize in the $2.50 to $3.00 range by Q4 2026.
Semiconductor Sector Diverges: Memory Soars, Nvidia Falls Behind
The divergence data is striking. Nvidia is up about 12% year-to-date 2026 but down about 3% over the past month. In the same period, the SMH semiconductor ETF is up 84% year-to-date and 15% over the past month. Micron has surged nearly 60% over the past month, with its stock price hitting an all-time high of about $1,089, cumulative gains of over 700% year-to-date, and a market cap exceeding $1.2 trillion. SanDisk has also risen nearly 60% over the past month, with a 52-week gain of over 4,400%. The market may not be bearish on AI—it just believes the AI value chain bottleneck is shifting. The previous logic was 'GPU scarcity → Nvidia pricing power → upstream wins.' The new logic is: GPU supply is loosening, but AI models' demand for high-bandwidth memory (HBM) and storage is exploding, making memory the new bottleneck.
Micron's latest quarterly report (Q2 2026) showed revenue of $23.8 billion, nearly triple the $8 billion in the same period last year. SanDisk, after its spin-off from Western Digital, reported Q3 FY2026 revenue of $5.95 billion, up 97% year-over-year. TrendForce data released on June 16 shows memory contract prices surged over 100% in the first half of 2026, with structural shortages expected to continue into the second half. Apple CEO Tim Cook admitted in an interview last week that Apple can no longer absorb the rising memory costs. When even Apple—a buyer with arguably the strongest bargaining power—publicly says it can't take it anymore, the pricing power of memory makers is evident. Micron will report its third-quarter results after the market closes tomorrow (June 24), with expectations of another record. This earnings report will be a key test of whether the 'memory super-cycle' can continue.
Goldman Sachs Trading Desk Chief: The Key Indicator Is Rental Price
Rich Privorotsky, head of Goldman Sachs' One-Delta trading desk, last week proposed a clear framework: if compute resources are truly scarce, rental prices should remain firm, supporting sustained capital expenditure. If supply increases and rental prices continue to decline, the 'compute shortage' assumption underpinning the entire AI hardware chain's valuation would be undermined. He further noted that this pressure would first be felt in hardware. The real beneficiaries are companies that sell complete systems and monetize through usage, not those selling only 'picks and shovels' upstream. The greater risk lies in the upstream hardware and infrastructure stack, where valuations are still built on the premise of 'persistent scarcity.' A recent 'Tokenomics' report from Citadel Securities echoed similar views: the core constraint for AI adoption has shifted from 'model capability' to 'cost and compute scarcity,' with users rapidly migrating to cheaper models. The Token price index fell for seven consecutive days, the longest losing streak this year. Santa Clara University finance professor Seoyoung Kim put it more bluntly: most buyers don't know how much compute they'll need next year, suppliers don't know how many GPUs to order, and Nvidia doesn't know how many to produce. All three are guessing, and when the direction of guessing simultaneously shifts from 'not enough' to 'maybe too much,' prices come under pressure.
SpaceX-Google $30 Billion Contract: The Long-Term Market Remains Hot
While spot rental prices are falling, the long-term contract market tells a different story. According to a SpaceX filing with the SEC on June 5, Google agreed to pay SpaceX $920 million per month from October 2026 to June 2029 to lease approximately 110,000 Nvidia GPUs along with supporting processors, memory, and other components. The total contract value is about $30 billion. Previously in May, Anthropic signed a similar agreement with SpaceX, paying $1.25 billion per month for all available compute power at its Colossus 1 data center in Memphis, with a total value of nearly $45 billion. These deals come after SpaceX completed its merger with xAI in February 2026, converting xAI's self-built Colossus supercomputing cluster into a commercial asset for leasing, locking in substantial revenue ahead of its IPO (target valuation $1.75 trillion).
For Nvidia, this is a mixed signal. On one hand, a long-term contract for 110,000 GPUs proves that large customers are still locking in massive compute capacity. RBC Capital Markets said after the deal that Nvidia is in the 'best position among peers' and that these GPU leasing agreements could at least in the short term allay market concerns about ASICs eating away at Nvidia's share. On the other hand, Google needs to lease from SpaceX precisely because its own built capacity cannot keep up with demand. Google's 2026 capital expenditure is between $180 billion and $190 billion; the $920 million monthly payment to SpaceX is less than 6% of its annual budget, essentially a 'bridge capacity.' When these hyperscalers' own data centers come online in 2027-2028, it remains questionable whether external leasing demand can sustain current levels. The contract also includes an early termination clause with a 90-day notice period. This doesn't look like a clause signed under conditions of 'extreme compute scarcity'; it looks more like a buyer keeping an exit option.
Nvidia's Risk: Not in Demand but in Pricing Power
Putting the pieces together, the challenge for Nvidia is that profit distribution within the AI value chain is shifting. On the GPU supply side, TSMC's yield improvements, more players securing inventory, and the impending mass availability of the B300 are three factors easing the extreme shortages of 2024-2025. On the demand side, hyperscale customers are still buying in large volumes, but the procurement pattern has shifted from 'buy at any cost' to 'price comparison, long-term contract locking, and retention of exit rights.' On the profit side, rental prices for downstream cloud providers are already falling. If Nvidia cannot simultaneously lower its chip prices, the margin squeeze in the middle will eventually eat into order volumes. The rise of memory chips as a new favorite is the other side of the value chain migration. The larger the AI model and the more inference tasks, the more rigid the demand for high-bandwidth memory. GPUs can improve efficiency through architectural upgrades, but memory bandwidth is a physical bottleneck with no shortcuts. Micron's HBM capacity for the entire year 2026 is already sold out. This 'unobtainable even with money' status stands in stark contrast to the falling rental price of Nvidia's B200. Micron's earnings report tomorrow will provide the next key data point. If revenue and guidance again beat expectations, the narrative that 'AI value chain is shifting from GPU to memory' will be further reinforced. For investors, this is not about being bearish on AI; it's about rethinking who on the AI chain has strengthening pricing power and whose is weakening.