Cloud Pricing Goes Dynamic

Cloud prices are moving again, and the movement looks like a change in the pricing regime rather than a run of one-off adjustments. Between September 17 and October 1, Nebius told customers that on-demand GPU rates would rise roughly 17 to 21 percent and memory-backed capacity about 41 percent, effective October 1.1 2 OVHcloud unbundled local storage and public IPv4 addresses from its Public Cloud instances on the same date, which moved the price of an identical configuration up 1.4 to 21.9 percent depending on size.3 CoreWeave announced that it has continued to contract new compute capacity at higher prices, on short terms of roughly three to six months.4 Under all three announcements sits the same pressure: memory. TrendForce’s September 30 survey projects conventional DRAM contract prices up another 10 to 15 percent in the fourth quarter, after a year of much larger increases, with server DRAM undersupplied into 2027.5

Two forces are repricing cloud at once: the cost of the components inside every instance, led by memory, and demand for AI capacity that keeps outrunning supply. Together they are moving prices at market frequency, between renewals, sometimes inside a single quarter. The question worth asking is what still shields a buyer. The honest answer: agreements help, they shield less than they used to, and market-driven pricing of the kind the book’s last chapter describes is arriving whether or not any of us opt in.

What changed in the last month

Nebius notified customers on September 17 rather than issuing a press release: new on-demand rates take effect October 1.1 The increases step up by chip generation, about 17 percent on H100 instances, 20 percent on H200, 19 percent on B200, 21 percent on the newest B300s, with AMD CPU rates up about 25 percent and memory-backed instances up roughly 41 percent, the largest single line in the notice. It was the company’s second increase in roughly three months.2 The stock market read the notice as pricing power rather than distress: Nebius shares climbed about 10 percent that morning, and operators that had sent no note at all caught a bid alongside it.6 The detail to notice is the memory line. Providers bundle memory into instance pricing, so it rarely appears on its own on a bill. A 41 percent increase on the memory line says the component inside every instance is repricing, not just the scarce GPUs on top.

OVHcloud did not raise an instance rate at all. What changed on October 1 is what the price includes. Local storage and public IPv4 addresses, previously bundled with Generation 3 instances, became separate billing lines, about €109 per terabyte-month for storage and about €1.97 per month for an address, and the total for an identical configuration rose 1.4 to 21.9 percent: a b3-8 went from €37 to €45 a month, a c3-4 from €33 to €41.3 The company also replaced its Savings Plans grid, cutting terms to 12 or 36 months at 15 and 30 percent discounts, where the old grid had offered 35 percent on a 12-month commitment. Customers already inside a plan keep their guaranteed compute rate until the end of the commitment, but from October 1 the storage and address lines bill at full price, outside the guarantee.

CoreWeave’s announcement was about the direction of new contracts: since June 30 it has signed new compute capacity at higher prices, and the quarter’s new contracts are short-dated, roughly three to six months, at about $40 million per megawatt on an annualized basis. Contracted power rose from about 3.7 gigawatts at June 30 to about 4.2 gigawatts by August 11. 4 Customers with billions in new commitments are signing short. When a seller offers shorter terms at higher prices, the market is saying that the length of the agreement, the certainty, is itself a product whose price went up.

None of these moves is unique to the last month; AWS raised EC2 Capacity Block rates 15 percent in January.7 What the last month adds is simultaneity: a neocloud, a European hoster, and the largest dedicated GPU lessor repricing within about two weeks of one another, with the same commodity underneath all three.

Why memory is driving it

The mechanism is production allocation. Making high-bandwidth memory for AI accelerators consumes DRAM fab capacity that previously made commodity and server memory, and only three companies, Samsung, SK hynix, and Micron, supply DRAM at scale, so the shift shows up quickly in the conventional market.8 TrendForce’s July survey put server DRAM contract prices up 13 to 18 percent in the third quarter alone and anticipated a server DRAM shortage in 2027; RDIMM bit supply is growing 15 to 20 percent a year while server CPU shipments grow faster, so buyers are building inventory ahead of the gap.9 JPMorgan’s research group estimates DRAM prices will have risen more than 400 percent from the start of 2024 to the end of 2026.8 Every cloud instance contains DRAM and every storage tier contains NAND flash, which TrendForce projects up 15 to 20 percent in the fourth quarter as well,5 so component repricing has a direct path into service pricing.

Demand is the second force, and it is not only GPUs. TrendForce notes cloud providers increasing procurement of server memory for general servers supporting agentic AI workloads.5 Nebius cited demand for AI compute exceeding available supply as the reason for its increase,1 and CoreWeave’s higher-priced, shorter-dated contracts are the same signal from the seller’s side. When demand outruns supply, a provider does not need a cost increase to reprice; the cost increase just removes the last restraint.

What an agreement buys, and what it no longer does

The standard way to get in front of repricing is to sign. The evidence says this works, partially. Several U.S.-based cloud providers have entered multi-year agreements with memory suppliers that restrict price increases for those clients; TrendForce’s July headline was that long-term agreements cap price increases,9 and its September survey expects some suppliers’ increases to lag the market because of LTA pricing mechanisms.5 At the service level, OVHcloud’s existing Savings Plan customers keep their guaranteed compute rate through the end of the commitment.3

The leaks are specific, and each one is visible in the last month’s announcements. The first is growth: TrendForce expects the primary source of server DRAM price increases to shift toward customers without long-term agreements, and toward incremental supply sold outside LTAs to existing LTA customers.9 The contracted base holds; the next tranche of growth pays the market price. An agreement shields what you already promised to buy, not what you will need next year. The second is the envelope: OVHcloud kept the guarantee and moved its boundary, with the compute line guaranteed and the storage and address lines that used to sit inside the bundle now billing separately, outside the discount.3 A guarantee that covers less of the bill each cycle is a weaker guarantee even when its percentage never moves. The third is the term: CoreWeave’s new contracts run three to six months at higher prices. 4 The length and rigidity of an agreement are things you buy, and their price moves with scarcity (chapter 20).

So the negotiation advice is not “sign everything” and not “sign nothing.” Sign, and read the agreement as a partial shield. Name what sits outside the guaranteed envelope (incremental supply, bundled components, renewal terms) and budget the parts that reprice as though they will reprice, because the last month says they do. Even a fully contracted enterprise watches part of its bill track the market: on-demand rates are market rates by definition, spot is priced by the provider and moves without bidding, exactly as it has since AWS replaced customer bidding with provider-set prices in 2017 (chapter 24), capacity blocks reprice between purchases, and the components inside every committed instance reprice continuously. A provider whose memory costs keep climbing cannot hold every bundled rate forever without unbundling something, which is exactly the move OVHcloud made. The share of the bill that reprices at market frequency is growing, and it is worth knowing which part of your bill that is.

The market the last chapter describes

The book’s closing chapter looks outward at exactly this drift (chapter 24). Compute already has markets; what 2026 added was an attempt to list futures on the rental price of a GPU. None of those contracts trades on a U.S. exchange yet: the CFTC extended its review of the NYMEX compute futures to November 9, citing novel or complex issues,10 and the ICE, Nodal, and Architect proposals are earlier-stage (chapter 24).

Why expect pricing to keep moving in this direction? Because the inputs are physically limited, memory fabrication, power, land, energy, and demand is still growing into those limits. When a resource is constrained and demand outruns it, prices are the mechanism other systems use to ration access; electricity markets run capacity auctions, and grain and freight price openly for the same reason. What cloud capacity has lacked is a shared signal of what it costs, because every rate is negotiated privately and kept confidential (chapter 24). A visible market price would not make shortages disappear. It would make the repricing everyone is already living through legible, comparable, and in time hedgeable.

Meanwhile, four moves need no exchange. Negotiate with the leaks in mind: audit what the agreement guarantees (rate, term, volume) and what sits outside the guarantee, and apply the price-of-certainty discipline to every clause (chapter 20). Track your basis: keep a monthly comparison of your actual rates against outside references; Silicon Data publishes H100 rental indexes, the nearest thing to a public reference today (chapter 24). Model renewals at market: treat next year’s renewal as repriced until proven otherwise; OVHcloud’s ladder, 35 percent to 15 percent on a 12-month term, is what that looks like.3 And watch November 9: if compute futures list, the first questions are the chapter’s four for Finance and Treasury, how closely the index tracks your rate, how much trades in the months you would hedge, who has authority to trade, and who funds the margin (chapter 24).

Nothing here says prices only rise; some of what is repricing now will reprice down when supply catches up. What changes is the regime: pricing that adjusts at market frequency, from many providers at once, instead of annually from a rate card. The teams that measure what they consume, keep outside references, and decide deliberately will read that repricing as information about scarcity rather than as an ambush.

Sources

1. Yahoo Finance, September 17, 2026: Nebius Group (NBIS) Raises AI Cloud Prices As Compute Demand Tests Customer Costs — customer communication rather than a public release; higher on-demand rates effective October 1, 2026; compute demand as the stated driver.

2. KuCoin News, September 2026: Nebius to Increase GPU Rental Prices by Up to 21% Starting October 1, and Heinrich Vogel, Shattered, updated September 23, 2026: Nebius Hikes AI Cloud Prices 21%, Memory Up 41% — per-tier figures compiled from the customer notice: H100 about 17 percent ($3.85 to $4.50 per GPU-hour), H200 about 20, B200 about 19, B300 about 21; AMD EPYC CPU about 25; memory-backed instances about 41; the second increase in roughly three months.

3. Octave Klaba, OVHcloud blog, October 1, 2026: Public Cloud: Pricing changes as of October 1, 2026 — identical configurations up 1.4 to 21.9 percent; hourly instance prices unchanged; local storage about €109 per terabyte-month and IPv4 about €1.97 per month as separate lines; b3-8 €37 to €45 and c3-4 €33 to €41 examples; Savings Plans reduced to 12- and 36-month terms at −15 and −30 percent effective September 1, 2026; existing plans guaranteed through the commitment with storage and IPv4 outside the guarantee from October 1.

4. CoreWeave press release, September 2026: CoreWeave Continues to Contract New Compute Capacity at Higher Prices — new capacity contracted at higher prices since June 30, 2026; third-quarter contracts short-dated at roughly three to six months and about $40 million per megawatt annualized; more than $25 billion of net new customer commitments early in the third quarter; contracted power from about 3.7 gigawatts at June 30 to about 4.2 gigawatts at August 11.

5. TrendForce, September 30, 2026: AI Server Demand Sustains Memory Contract Price Increases in 4Q26, While Consumer-Side Pressure Persists — conventional DRAM contract prices up 10–15 percent quarter-over-quarter and NAND flash up 15–20 percent in 4Q26; server DRAM undersupplied in 4Q26; cloud providers increasing procurement for general servers supporting agentic AI workloads; some suppliers’ increases lagging the market average due to pricing mechanisms in long-term agreements.

6. David Moadel, 24/7 Wall St., September 17, 2026: IREN Climbs 6%, Nebius Jumps 10% as GPU Rental Rates Head Higher; CoreWeave Slips — Nebius notified customers of higher on-demand rates effective October 1; Nebius shares up about 10 percent that morning; read-through repricing across neocloud operators, including a bid for operators with no announcement of their own.

7. InfoQ, January 2026: AWS Hikes EC2 Capacity Block Rates by 15% in Uniform ML Pricing Adjustment — a 15 percent increase in EC2 Capacity Block rates for ML instances, as context that 2026 repricing predates the last month.

8. JPMorgan Global Research: The AI-Driven Memory Shortage: DRAM Prices, Inflation and Market Risks — DRAM prices estimated to rise more than 400 percent from the start of 2024 to the end of 2026; and Network World: What’s causing the memory shortage? — Micron, SK hynix, and Samsung as the three core DRAM suppliers.

9. TrendForce, July 9, 2026: Long-Term Agreements Cap Price Increases; Server DRAM Contract Prices Expected to Rise 13–18% QoQ in 3Q26 — multi-year long-term agreements with U.S.-based cloud providers restricting supplier price increases for those clients; server DRAM up 13–18 percent quarter-over-quarter in 3Q26; RDIMM bit supply growing 15–20 percent year-over-year against faster server CPU growth; a server DRAM shortage anticipated for 2027; increases from the third quarter shifting toward customers without LTAs and toward incremental supply sold outside LTAs to existing LTA customers.

10. CFTC, Division of Market Oversight letter to CME Group, September 21, 2026 (product record 62544) — extension of the review of NYMEX Submission No. 26-370, the compute futures CME Group and Silicon Data announced on August 11, 2026 with trading to start October 5, to November 9, 2026, citing “novel or complex issues.” As of this posting, no compute futures contract trades on a U.S. exchange.