Reserved GPU Capacity Bidding: Capacity Blocks and MSAs (2026)

Reserved GPU capacity bidding in 2026 isn't one market with one price. It's three separate mechanics that happen to solve the same problem, plus a contract vehicle that sits underneath the largest deals, and most buyers only ever see one of them before assuming it's the whole picture.

AWS Capacity Blocks for ML just got more expensive for the third time this year, effective October 7, 2026, according to explainx.ai's coverage of the hike. That's the number most buyers check first because AWS publishes it openly, but it's an administratively set rate card, not a bid you can beat. SF Compute runs something closer to an actual market. Compute Exchange runs its own version of a live auction. And the biggest reservations, the ones measured in years and tens of millions of dollars, don't go through any of those paths at all; they go through a master service agreement. This post maps all four pieces against the fresh AWS rate card and where Spheron's own quote-based reservation path fits next to them, stopping short of the clause-by-clause contract negotiation that a companion guide already covers in depth.

TL;DR: How Reserved GPU Capacity Bidding Actually Works in 2026

Reserved GPU capacity bidding in 2026 isn't one auction, it's three different acquisition mechanics plus the MSA contract vehicle that backs the largest deals.

For the full rate card across every Capacity Block instance family, see GPU cloud pricing compared across providers.

What "Reserved GPU Capacity Bidding" Means (There Is No Single Auction)

Search for "reserved GPU capacity bidding" and you'll find pages that treat it as one thing: a market where buyers submit bids and the highest one wins a block of GPUs for a fixed term. That market exists in pieces, but it isn't unified, and conflating the pieces is how teams end up budgeting against the wrong mechanic.

Three structurally different models are live right now. AWS sells Capacity Blocks at a price it sets itself and republishes periodically, with zero room to negotiate down. SF Compute lets tenants who've already committed to reserved capacity resell their idle time to other tenants above a floor price, which is the closest thing to real price discovery in this space (more on how that compares to Spheron's own marketplace later in this post). Compute Exchange is the one that actually calls itself an auction: buyers submit a GPU type, quantity, region, and duration, and the platform returns real-time pricing from verified providers as a side-by-side comparison in under two hours. None of the three is the kind of continuously open, anyone-can-watch order book that "bidding" usually implies. "Bidding" is the shorthand the market has landed on for "a reserved-capacity acquisition process with some element of price competition or negotiation," and it covers mechanics that behave very differently once you're the one signing the term sheet.

Understanding which model you're actually in matters because the GPU supply picture underneath all three is the same: tight. The 2026 GPU shortage is narrower than it was in 2025, but it's still real at the top end, and scarcity is exactly what gives AWS the room to raise its Capacity Block rate card three times in ten months without losing buyers to a competing published price.

AWS Capacity Blocks Pricing: The Benchmark Floor Everyone Measures Against

AWS Capacity Blocks for ML are the reference point every other reserved-GPU mechanic gets compared against, because AWS publishes the number and nobody else at this scale does. As of the October 7, 2026 update, the rate per accelerator-hour in US regions is:

Source: AWS's own Capacity Blocks pricing page. GovCloud runs higher on the two Blackwell families: $16.819/accelerator-hr for P6-B300 and $14.801/accelerator-hr for P6-B200. A p5e.48xlarge block, at $6.866 per accelerator times 8 GPUs, comes out to roughly $54.93/hr for the full node, the exact math behind the FAQ question teams keep asking about that specific instance type.

AWS list prices change without notice. The rates above reflect AWS's published Capacity Blocks pricing page as of 7 Oct 2026 and may have changed since. Check aws.amazon.com/ec2/capacityblocks/pricing for current figures.

This is the third time this rate card has moved in 2026. If you read our earlier breakdown of the January and July hikes when they landed, treat every number in it as superseded: the $14.04 P6-B300 figure there was accurate as of July 15, 2026, and the $16.146 rate in the table above is what replaced it after October 7.

The October 7, 2026 Hike: Third Increase This Year

AWS raised Capacity Block prices approximately 15% across all regions effective October 7, 2026, with non-US regions, notably P5 and P5en, seeing increases up to 26.5% as AWS moved to uniform global pricing and eliminated the regional discounts those instance families used to carry, according to explainx.ai. That's the third Capacity Block price increase of 2026 on the same product line: roughly 15% on January 4, roughly 20% on July 1, and roughly 15% again on October 7, per Network World's reporting, which puts the cumulative increase on P6-B300, P6-B200, and the P5 family since January at roughly 58%. The same report traces the P5e line through the first of those hikes alone: p5e.48xlarge ran $34.608/hr before January 4 and $39.799/hr after it in most US regions, with US West (N. California) landing at $49.75/hr. Run that same instance through July's ~20% increase and October's ~15% increase on top, and the current $6.866 per-accelerator rate (8 x $6.866 = roughly $54.93/hr) is three compounding hikes on one instance family in ten months.

Full Rate Card: What Changed and What Didn't

Not every family has been hit on every round. P6-B300 and P6-B200 only entered the Capacity Block catalog partway through the year, so they've absorbed one fewer increase than P5e and P5en, which have now been through all three. That unevenness matters if you're comparing a quote from six months ago against today's rate card: the percentage move since your last quote depends on which instance family you're actually buying, not a single blanket number across the whole product. A Capacity Block is not the same product as an AWS Reserved Instance or Savings Plan, either. A Reserved Instance discounts a flexible hourly rate over a one- or three-year term with no guarantee the hardware is physically available when you want to launch. A Capacity Block reserves a specific block of GPUs for a fixed, non-cancellable window while paying only for the amount of time that you need, according to AWS's own documentation; you're buying a hardware guarantee, not a discount, which is exactly why it carries a premium over on-demand rather than a cut.

Three Live Models for Buying Reserved GPU Capacity

Once you get past the AWS rate card, the reserved-capacity market splits into three genuinely different buying experiences. None of them require you to beat a published floor price the way a real auction would; each trades a different amount of price flexibility for a different amount of speed.

AWS Capacity Blocks: Fixed Price, Fixed Window, No Negotiation

This is the simplest of the three because there's nothing to negotiate. You pick an instance family, a GPU count, and a start date from AWS's console, and you pay whatever the current rate card says for that fixed, non-interruptible window. There's no form of back-and-forth, no quote request, and no discount for buying more: the per-accelerator rate is the same whether you reserve one GPU or a full cluster's worth. The tradeoff for that simplicity is that you're accepting AWS's number on AWS's schedule, including whatever the next rate change brings.

SF Compute and Compute Exchange: Reserved-Price and Aggregated-Quote Markets

For a full side-by-side on how SF Compute's cluster-focused, CLI-driven buying experience compares with a broader on-demand marketplace, see our Spheron vs SF Compute breakdown.

Compute Exchange works differently again, and it's the only one of the three that uses the word auction about itself. Reservation terms run 1, 3, 6, 12, 24, or 36 months. It's the closest thing in this lineup to literal price-discovery bidding, run across a network of neocloud partners rather than against a single supplier's own rate card, even though a two-hour comparison window still isn't a continuously open order book.

Spheron's Quote Path: No Bidding, One Negotiated Price in 24-48 Hours

Spheron's reserved-commitment path skips bidding entirely and runs as a four-step quote process instead, per Spheron's reserved commitments page. You submit the GPU model, quantity, duration, interconnect preferences, and target start date; the team checks availability across certified data center partners and returns a single quote within 24-48 hours; you review a proposal covering hardware specs, pricing, location, and terms; and once you confirm, provisioning begins. Reservations run from 8 GPUs up to 512 or more, for terms from one week to 12 or more months, with volume pricing that improves as the GPU count and term length grow. That 24-48 hour turnaround beats the back-and-forth of a multi-vendor RFQ, and it's a wider window than Compute Exchange's roughly two-hour comparison, trading speed for a single negotiated number instead of a side-by-side list of competing quotes.

This path has real limits worth being honest about. There's no self-serve bidding, auction, or published floor-price rate card the way AWS publishes its Capacity Block list; a buyer who wants to see competing numbers side by side before committing has to request the quote first rather than browsing a live price table. There's also no in-app contract e-signing flow described on the page; large or long-term reservations get negotiated and documented the way MSA-based neocloud deals generally are, not through an automated in-app signature step. And the reservation path itself starts at 8 GPUs, so a team that needs one to seven GPUs on a fixed term is better served by Spheron's on-demand or spot catalog than by the reservation form. Documentation for provisioning once a reservation is confirmed is in the Spheron docs.

MSAs: The Contract Vehicle Behind Large Reserved GPU Deals

None of the three models above is how the biggest reserved-GPU deals actually get signed. A multi-year, tens-of-millions-of-dollars commitment doesn't go through a self-serve console purchase, a resale marketplace, or even a quote form; it goes through a master service agreement, usually paired with one or more order forms or statements of work.

What an MSA Plus Order Form Actually Covers

An MSA is the standing legal contract between buyer and supplier: liability caps, intellectual property terms, security and compliance commitments, data handling, and the SLA framework that applies across every reservation signed under it. It doesn't by itself specify a GPU model or a price. That detail lives in an attached order form or statement of work, a shorter document that references the MSA's standing terms and fills in the specifics for one particular reservation: which GPUs, how many, for how long, at what rate, starting when. The structure lets a buyer and supplier negotiate the hard legal terms once and then add new capacity commitments over time without renegotiating the whole contract from scratch.

The line items inside that order form, overage rates if you exceed your committed capacity, step-down clauses if your usage drops, exit and early-termination terms, are exactly what our guide to negotiating a GPU cluster reservation contract walks through in detail. This post stops at the contract vehicle itself; that one covers the clause-by-clause negotiation.

A Real Example: A 60-Month GPU MSA and What It Committed To

In September 2026, Quanome Technologies' subsidiary XDT Infrastructure signed a Token Factory Master Services Agreement for AI inference computing capacity on NVIDIA-accelerated infrastructure, according to the company's SEC 8-K filing. The deal covers a dedicated, physically isolated GPU pool under a 60-month commitment term, with aggregate committed fees of approximately $100.9 million over that period.

That scale and duration is what an MSA is built for and what none of the other three models above are designed to handle. AWS Capacity Blocks cap out at the instance families and durations on a published catalog page; SF Compute's resale market and Compute Exchange's aggregated quotes both operate well below a 60-month, nine-figure commitment in practice. An MSA with an attached order form is the vehicle large buyers and suppliers reach for specifically when the deal is too big, too long, or too custom to fit a self-serve purchase flow.

Reserved Bidding vs On-Demand vs Spot: Which Fits Your Workload

None of the reserved-capacity mechanics above are the right call for every workload. The decision mostly comes down to how predictable your GPU usage is and how far in advance you can commit to it.

• Reserved (any of the four models above) fits workloads with a known, steady GPU floor over months, a production inference service with predictable traffic, or a training run with a locked schedule where losing the hardware mid-run isn't an option.
• On-demand fits workloads with variable or unpredictable usage, or teams that haven't yet nailed down how much capacity they'll need month over month. Our guide to planning and sourcing GPU capacity for AI deployment walks through building that usage picture before you commit to any reservation.
• Spot fits checkpointable, interruption-tolerant jobs, batch training, evaluation runs, async pipelines, where the savings outweigh the risk of a mid-run preemption. Our deep dive on GPU spot instance arbitrage, bidding, and failover strategy covers how fixed-bid, dynamic-bid, and market-price strategies actually work for the one corner of this market that does behave like a real-time auction.

A reasonable default for a growing team: run your steady-state floor on a reservation, whichever of the four mechanics above fits your scale and timeline, and burst above that floor on spot or on-demand rather than overcommitting to a reservation sized for a peak you'll rarely hit.

Need a reserved GPU block without bidding against a rate card or negotiating a multi-year MSA for a mid-size commitment? Spheron's reserved-commitment path returns a single quote in 24-48 hours for 8 to 512+ GPUs, from a one-week test run to a 12+ month commitment.

Compare live GPU pricing → | Get started on Spheron →