GPU rental prices have dropped 30% over three weeks, and the AI value chain is shifting from NVIDIA's "Great Migration" to memory chips.

Author: Claude, Deep Tide TechFlow

Deep Tide Guide: Nvidia's B200 chip leasing price has fallen from a high of $6.11/hour at the end of May to $4.22/hour, a drop of about 30% in three weeks. Meanwhile, the semiconductor sector has experienced rare divergence: the SMH semiconductor ETF has risen 15% over the past month, Micron and SanDisk have surged nearly 60%, while Nvidia has fallen 3%. For those holding Nvidia or considering AI infrastructure investments, a key question has emerged: the money in AI hasn't decreased; it has just moved elsewhere.

Nvidia still gained about 12% this year, but current market attention seems to have shifted away from it.

Over the past month, VanEck Semiconductor ETF (SMH) surged 15%, Micron and SanDisk each soared nearly 60%. Nvidia not only lagged behind but also declined about 3 against the trend. More tellingly, the core indicator supporting Nvidia's valuation narrative—the cloud leasing price of the B200 chip—has also weakened.

According to GPU compute pricing platform Ornn data, the hourly leasing price of the B200 reached a three-month high of $6.11 on May 30, then continued to decline, falling to $4.22 by the end of last week, a decrease of about 30%. Goldman Sachs' One-Delta trading desk head Rich Privorotsky directly addressed this last week: the myth of "compute scarcity" in AI may be falling from the altar.

B200 leasing prices have fallen 30% in three weeks, putting pressure on the "compute scarcity" narrative

Nvidia's B200 is the core compute chip for current large-scale data centers, and its leasing price is seen as a barometer of supply and demand for AI infrastructure. Data from multiple third-party tracking platforms shows that B200 pricing is softening.

Ornn data shows that the B200 hourly leasing price has been declining from the $6.11 high on May 30 to $4.22 last weekend. AIMultiple's compilation of 63 cloud service providers' monthly price indices shows the median B200 quote at $6.11/hour, but new cloud providers (neocloud) have already pushed their bottom prices down to $3.44. GetDeploying's data on 26 B200 cloud providers is even more extreme: an average of $4.99/hour, with the lowest quotes at just $2.25/hour (for three-year reserved contracts).

Three factors are driving the price decline: TSMC's improved yield on 4N process reduces B200 production costs; SK Hynix and Micron's HBM3e supply will significantly loosen in Q2 2026; and more new cloud vendors have obtained B200 inventory, with companies like RunPod, Lambda, Nebius, and Spheron already offering spot stock, which has driven prices down overall.

Pressure will intensify in the second half of the year. After Nvidia's next-generation Blackwell Ultra B300 enters the spot market, some B200 capacity will shift from on-demand to bidding mode. Spot prices for B300 have already dropped as low as $2.45/hour, even cheaper than the minimum listing price for B200. Institutions like Spheron and Thunder Compute predict that B200 on-demand prices may stabilize between $2.50 and $3.00 in Q4 2026.

For Nvidia investors holding positions, the softening leasing prices mean profit margins for downstream customers (cloud providers, new cloud platforms) are under pressure, and their purchasing willingness directly influences Nvidia's order flow.

Semiconductor sector divergence: memory surges, Nvidia falls behind

This divergence is quite striking.

Nvidia has gained about 12% so far in 2026, but has fallen about 3% over the past month. Meanwhile, the SMH semiconductor ETF has risen 84% year-to-date, with a 15% increase in the past month. Micron has surged nearly 60% in the past month, hitting a record high of around $1,089, with a year-to-date increase of over 700%, and a market cap surpassing $1.2 trillion. SanDisk has also risen nearly 60% in the past month, with a 4,400% increase over the past 52 weeks.

The market's attitude may not be that AI is no longer promising, but that the bottleneck in the AI value chain is shifting.

The previous logic was "GPU scarcity → Nvidia has pricing power → upstream profits are highest." Now, the narrative has shifted: GPU supply is easing, but demand for high-bandwidth memory (HBM) and storage in AI models is exploding, making memory the new bottleneck.

Micron's latest quarterly report (Q2 2026) shows revenue of $23.8 billion, nearly doubling year-over-year (from $8 billion last year); SanDisk, spun off from Western Digital, reported revenue of $5.95 billion in Q3 2026, up 97% year-over-year.

Data from TrendForce released on June 16 indicates that memory contract prices soared over 100% in the first half of 2026, and structural shortages are expected to persist into the second half. Apple CEO Tim Cook admitted last week in an interview that Apple can no longer absorb the rising memory costs. When even Apple, with its strongest bargaining power, publicly states "it can't handle it," the pricing power of memory manufacturers is evident.

Micron will release its Q3 2026 earnings after market close tomorrow (June 24), with expectations of another record. This report will be a key indicator of whether the "memory supercycle" can continue.

Goldman Sachs trader: core indicator is leasing price

Rich Privorotsky, head of Goldman Sachs' One-Delta trading desk, laid out a clear framework last week:

If compute resources are truly scarce, leasing prices should remain firm, and sustained capital expenditure is justified. If supply increases and leasing prices continue to decline, the core assumption of "compute scarcity" supporting the valuation of the entire AI hardware chain will be shaken.

He further pointed out that this pressure will first manifest on the hardware side. The real beneficiaries are companies that sell complete systems and monetize through usage, rather than upstream suppliers selling "picks and shovels." The greater risk lies in the upstream segments of hardware and infrastructure, where valuations are still based on the premise of "persistent shortages."

The message is clear: Nvidia's business model is selling chips (picks and shovels), not charging based on usage. If downstream customers' leasing prices fall but Nvidia's chip prices do not, profit margins will be squeezed, ultimately leading to order slowdowns.

Citadel Securities' recent "Tokenomics" report echoes this view: the core constraints of AI adoption have shifted from "model capability" to "cost and compute scarcity," with users accelerating migration to cheaper models. The token price index has fallen for seven consecutive days, marking the longest decline this year.

Seoyoung Kim, a finance professor at Santa Clara University, put it plainly: 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 much to produce. All three are guessing, and when expectations shift from "not enough" to "maybe too much," prices will come under pressure.

SpaceX-Google's $30 billion mega-contract: the spot market remains hot

While spot leasing prices are falling, the long-term contract market tells a different story.

According to SpaceX filings with the SEC on June 5, Google agreed to pay SpaceX $920 million per month from October 2026 to June 2029 to lease about 110k Nvidia GPUs along with supporting processors, memory, and other components. The total contract value is approximately $30 billion. In May, Anthropic signed a similar agreement with SpaceX, paying $1.25 billion per month to lease all available capacity at its Colossus 1 data center in Memphis, totaling nearly $45 billion.

These contracts are set against the backdrop of SpaceX's merger with xAI in February 2026, after which it converted its self-built Colossus supercomputing cluster into a commercial leasing asset, locking in substantial revenue ahead of its IPO (target valuation of $1.75 trillion).

For Nvidia, this is a contradictory signal. On one hand, the long-term contracts for 110k GPUs demonstrate that major clients are still locking in compute capacity at scale. RBC Capital Markets, after the announcement, said Nvidia "is in the best position among peers," believing these leasing agreements can at least temporarily dispel concerns about ASICs eating into Nvidia's market share.

On the other hand, Google needs to lease capacity from SpaceX precisely because its own capacity can't keep up with demand. Google's capital expenditure in 2026 is projected between $12k and $190 billion, and SpaceX's monthly payment of $920 million accounts for less than 6% of that annual budget—essentially a "bridge capacity." When these mega-clients' own data centers come online in 2027-2028, whether external leasing demand can sustain current levels remains uncertain.

The contracts also include a 90-day notice period for early termination. This is less like a contract signed during a time of "extreme compute scarcity" and more like a buyer leaving a back door open.

Nvidia's risk: not on the demand side, but in pricing power

Connecting these clues, Nvidia faces a shift in profit distribution along the AI value chain.

On the supply side, TSMC's improved yields, more vendors obtaining inventory, and the upcoming large-scale launch of B300 are easing the extreme shortages expected in 2024-2025. On the demand side, mega-clients are still purchasing at scale but are shifting from "buy at any cost" to "price comparison, long-term contracts, and retention of exit rights." Profits are under pressure as leasing prices for downstream cloud providers decline; if Nvidia's chip prices don't adjust downward in tandem, profit margins will be squeezed, ultimately impacting order volume.

The rise of memory chips as a new favorite reflects another aspect of the value chain shift.

As AI models grow larger and inference tasks increase, demand for high-bandwidth memory (HBM) becomes more rigid. While GPUs can improve efficiency through architectural upgrades (e.g., B200's FP4 precision halves per-parameter byte), memory bandwidth is a physical bottleneck with no shortcuts. Micron's HBM capacity was sold out for all of 2026, a state of "money can't buy" that starkly contrasts with the declining leasing prices of Nvidia's B200.

Tomorrow's earnings report from Micron will provide another key data point. If revenue and guidance again surpass expectations, the narrative of "AI value chain shifting from GPU to memory" will be further reinforced. For investors, this isn't about being bearish on AI but about reconsidering who has pricing power in this chain and who is losing it.