NVIDIA B200 GPU rental prices drop 30% in three weeks as the AI value chain shifts toward memory chips.

Author: Claude, Deep潮 TechFlow

Shenchao Digest: The rental price of NVIDIA's B200 chip has dropped from its late-May peak of $6.11 per hour to $4.22 per hour, a decline of approximately 30% over three weeks. Meanwhile, the semiconductor sector has shown rare divergence: the SMH semiconductor ETF has risen 15% over the past month, while Micron and SanDisk have each surged nearly 60%; in contrast, NVIDIA has fallen 3% during the same period. For those holding NVIDIA shares or considering investments in AI infrastructure, a key question has emerged: AI funding hasn't decreased—it's simply shifted elsewhere.

NVIDIA still rose about 12% this year, but market attention now seems to have moved away from it.

Over the past month, VanEck’s Semiconductor ETF (SMH) surged 15%, with Micron Technology and SanDisk each climbing nearly 60%. NVIDIA, however, failed to keep pace and instead declined by approximately 3%. More telling is that the core metric underpinning NVIDIA’s pricing narrative—the cloud rental price of the B200 chip—has also softened in tandem.

According to data from the GPU computing power pricing platform Ornn, the hourly rental price of the B200 reached a three-month high of $6.11 on May 30, but has since declined steadily, dropping to $4.22 by last weekend—a decrease of approximately 30%. Last week, Rich Privorotsky, Head of the One-Delta Trading Desk at Goldman Sachs, directly addressed the issue: the myth of AI “computing power scarcity” may be losing its luster.

The NVIDIA B200 is the core computing chip for today's hyperscale data centers, and its rental price is regarded as a barometer of supply and demand in AI infrastructure. Data from multiple third-party tracking platforms indicate that B200 pricing is beginning to soften.

Ornn data shows that the hourly rental price for the B200 dropped from a peak of $6.11 on May 30 to $4.22 by last weekend. According to a monthly price index compiled by AIMultiple for 63 cloud providers, the median quote for the B200 is $6.11 per hour, but new cloud (neocloud) vendors have already lowered their bottom prices to $3.44. Data from GetDeploying, tracking 26 B200 cloud providers, reveals even more extreme figures: an average price of $4.99 per hour, with the lowest quoted rate at just $2.25 per hour (for a three-year reserved contract).

Three factors are driving the price decline: TSMC’s improved 4NP process yield has reduced B200 production costs; HBM3e supply from SK Hynix and Micron is expected to significantly ease in Q2 2026; and more new cloud providers have acquired B200 inventory, with RunPod, Lambda, Nebius, Spheron, and others now offering spot inventory, increasing competition and pressuring overall prices.

Pressure will increase further in the second half of the year. Once NVIDIA’s next-generation Blackwell Ultra B300 enters the spot market, some B200 capacity will shift from on-demand to spot pricing. The spot price for B300 has already dropped as low as $2.45 per hour, cheaper than the lowest listed price for B200. Institutions such as Spheron and Thunder Compute predict that B200 on-demand prices could stabilize in the $2.50 to $3.00 range by Q4 2026.

For investors holding NVIDIA stock, a softening lease price indicates pressure on the profit margins of NVIDIA’s downstream customers (cloud providers and new cloud platforms), whose purchasing intent directly determines NVIDIA’s order pace.

The data from this round of divergence is quite striking.

NVIDIA has risen approximately 12% year-to-date in 2026 but declined about 3% over the past month. During the same period, the SMH semiconductor ETF has gained 84% this year and 15% over the past month. Micron Technology has surged nearly 60% over the past month, reaching a historic high of approximately $1,089 per share, with a year-to-date gain exceeding 700% and a market capitalization surpassing $1.2 trillion. SanDisk has also risen nearly 60% over the past month and more than 4,400% over the past 52 weeks.

The market may not be losing faith in AI; rather, it believes the bottleneck in AI’s value chain is shifting.

The previous logic was: "GPU scarcity → NVIDIA holds pricing power → upstream players earn the most." The current logic has shifted: GPU supply is easing, but demand for high-bandwidth memory (HBM) and storage from AI models is surging, making memory the new bottleneck.

Micron's latest quarterly earnings (Q2 2026) reported revenue of $23.8 billion, nearly tripling year-over-year (compared to $8 billion in the same period last year); after its spin-off from Western Digital, SanDisk reported revenue of $5.95 billion for Q3 2026, a 97% year-over-year increase.

According to data released by TrendForce on June 16, memory contract prices surged over 100% in the first half of 2026, and structural shortages are expected to persist into the second half. Apple CEO Tim Cook acknowledged last week in an interview that Apple can no longer absorb the rising costs of memory. When even Apple—the buyer with the strongest bargaining power—publicly admits it can’t keep up, the pricing power of memory manufacturers is evident.

Micron will report its third-quarter earnings after the market closes tomorrow, June 24, with widespread market expectations for another record performance. This earnings report will serve as a key test of whether the memory supercycle can continue.

Last week, Rich Privorotsky, head of the Goldman Sachs One-Delta trading desk, presented a clear framework for evaluation:

If computing power resources are truly scarce, rental prices should remain robust, justifying sustained capital expenditures. If supply increases and rental prices continue to decline, the core assumption underpinning the valuation of the entire AI hardware chain—computing power shortage—would be undermined.

He further noted that this pressure will first manifest at the hardware level. The true beneficiaries are companies that sell complete systems and monetize through usage, rather than upstream players selling only “picks and shovels.” The greater risk lies in the upstream segments of the hardware and infrastructure stack, where valuations still rely on the premise of “persistent scarcity.”

The intent of this statement is clear: NVIDIA’s business model is selling chips (the pickaxes and shovels), not charging based on usage. If downstream customers’ rental prices are falling but NVIDIA’s chip prices remain unchanged, this creates margin pressure that ultimately translates into slower order growth.

Citadel Securities' recent "Tokenomics" report also echoes this assessment: the primary constraint on AI adoption has shifted from "model capability" to "cost and computational scarcity," with users rapidly migrating to cheaper models. The token price index has fallen for seven consecutive days, marking the longest decline since the beginning of the year.

Seoyoung Kim, a finance professor at Santa Clara University, puts it more plainly: most buyers don’t know how much computing power 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 parties are guessing, and when their guesses shift simultaneously from “not enough” to “maybe too much,” prices come under pressure.

The spot rental price is falling, but the long-term contract market tells a different story.

According to a filing submitted by SpaceX to the SEC on June 5, Google has agreed to pay SpaceX $9.2 billion per month from October 2026 through June 2029 to lease approximately 110,000 NVIDIA GPUs along with accompanying processors, memory, and other components. The total value of the contract is approximately $30 billion. Previously in May, Anthropic signed a similar agreement with SpaceX, paying $12.5 billion per month to lease all available computing capacity at SpaceX’s Colossus 1 data center in Memphis, with a total value of nearly $45 billion.

The context for these two contracts is that, following SpaceX's merger with xAI in February 2026, SpaceX will convert xAI's previously self-built Colossus supercomputing cluster into a commercial asset available for rent, securing substantial revenue ahead of its IPO (target valuation of $1.75 trillion).

For NVIDIA, this is a mixed signal. On one hand, the long-term contract for 110,000 GPUs demonstrates that major clients continue to secure substantial computing power. RBC Capital Markets stated after the announcement that NVIDIA is “in the strongest position among its peers,” suggesting that these GPU leasing agreements at least temporarily alleviate market concerns about ASICs eroding NVIDIA’s market share.

On the other hand, Google needs to lease computing power from SpaceX precisely because its own production capacity cannot keep up with demand. Google’s capital expenditure for 2026 is projected to be between $180 billion and $190 billion; SpaceX’s monthly payment of $920 million represents less than 6% of that annual budget, essentially serving as “bridge capacity.” As these major clients’ own data centers come online between 2027 and 2028, it remains uncertain whether external leasing demand will sustain its current scale.

The contract also includes a termination clause with a 90-day notice period. This does not resemble a clause signed during a period of extreme hash rate scarcity, but rather appears to provide the buyer with an exit strategy.

Putting the above clues together, NVIDIA's challenge is that profit distribution in the AI value chain is shifting.

On the supply side, TSMC’s yield improvements, increased inventory from more manufacturers, and the upcoming large-scale release of the B300 are jointly alleviating the extreme shortages of 2024–2025. On the demand side, hyperscale customers continue to make large purchases, but their approach has shifted from “securing chips at any cost” to “comparing prices, locking in long-term volumes, and retaining exit options.” On the profit side, rental prices from downstream cloud providers are already declining; if NVIDIA does not simultaneously reduce its chip prices, margin pressure in the middle will ultimately erode order volumes.

The rise of memory chips as a new favorite is another side of the value chain shift.

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 (such as B200’s FP4 precision, which halves the bytes per parameter), but memory bandwidth remains a physical bottleneck with no shortcuts. Micron’s HBM capacity for the entire year of 2026 is already sold out, creating a “can’t buy it even if you have money” situation that stands in stark contrast to the declining rental prices of NVIDIA’s B200.

Micron’s earnings report tomorrow will provide the next key data point. If revenue and guidance once again exceed expectations, the narrative of “the AI value chain shifting from GPUs to memory” will be further reinforced. For investors, this is not about being bearish on AI—it’s about reconsidering which players along the AI value chain are gaining pricing power and which are losing it.