What happened to GPU prices this year?
For two years, the price of rented AI compute did one thing. It fell. H100 rates that spiked above $8 per GPU-hour in the 2023 shortage slid through 2024 as supply caught up, and in June 2025 AWS cut on-demand pricing for its H100 P5 instances by 44%, with comparable reductions following at Google Cloud and Azure. Budgeting for 2026 on the assumption that the slide would continue was the consensus move. It was also wrong.
The turn came in the fall of 2025, and it came in the contract market first. SemiAnalysis's one-year H100 rental price index, built from a monthly survey of more than 100 market participants and validated against transaction data, recorded $1.70 per GPU-hour in October 2025. By the end of March 2026 it stood at $2.35, up about 40%, because on-demand capacity had sold out across nearly every GPU generation and holders of locked-up capacity would not release it back into the pool.
Spot and on-demand series followed the contract market up, with one pause in the summer. Silicon Data's neocloud H100 index, $2.37 per GPU-hour when it launched in May 2025, troughed at $2.00 in December 2025 and reached $2.82 on October 10, 2026, and $2.88 by October 24. JPMorgan put the average H100 on-demand rental in the non-hyperscale market at $2.68 per GPU-hour in July 2026, down 1.1% on the month and the first decline after seven consecutive increases, with utilization still above 70%. GetDeploying's median on-demand H100 rate across 41 providers was $3.50 in October 2026, about 13% above a year earlier. The H200 median rose about 26% over the same twelve months, to $4.70.
Blackwell, the current generation, was the most volatile line in the market. Its rental index rose 27.6% year to date by September 2026, after a spring round trip that took spot rates from $2.75 in mid-February to $4.08 by mid-April on Ornn's live transaction index. By October 10, Silicon Data's indexes stood at $5.93 per GPU-hour for the B200 and $6.71 for the B300.
Into the fall, list prices followed the market up, which is the step that turns a squeeze into a reset. These are posted prices, the ones a new customer is quoted before any negotiation.
| Posted price increase | Was | Now | When |
|---|---|---|---|
| Lambda, H100 SXM on demand | $2.99 | $3.99 | June 2026 |
| Runpod Secure Cloud, H100 SXM | $3.29 | $3.99 | By October 2026 |
| Nebius, HGX H100 on demand | $3.85 | $4.50 | October 1, 2026 |
| Nebius, HGX B300 on demand | $7.85 | $9.50 | October 1, 2026 |
| AWS P5 Capacity Blocks, per H100 | $5.191 | $5.970 | August to October 2026 |
The older generation rising is the tell. The H100 is two product cycles behind the frontier now, and its price is going up anyway, because the binding constraint is not the chip. It is the powered, memory-equipped capacity around the chip.
Who charges what for the same silicon?
An H100 is an H100. The hourly price of renting one in October 2026 ranged from under $2 on peer-to-peer marketplaces to $12.29 at the top of the hyperscaler list stack, and the spread did not narrow during the year. It stratified, into a market layer where prices float and a list layer where they are posted, premium, and sticky.
| Eight-GPU H100 node, on-demand list | Per hour | Per GPU |
|---|---|---|
| CoreWeave, HGX node | $49.24 | $6.16 |
| AWS, p5.48xlarge | $55.04 | $6.88 |
| Google Cloud, a3-highgpu-8g | $88.49 | $11.06 |
| Microsoft Azure, ND96isr H100 v5 | $98.32 | $12.29 |
Below the hyperscalers, the specialist clouds price the same GPU by the hour at $3.99 at Lambda, $3.99 at Runpod, and $4.50 at Nebius after its October increase, against the $2.82 neocloud index and the $3.50 market median. What the hyperscaler premium buys is everything around the GPU: enterprise agreements, global regions, networking and storage bundles, support, and capacity that is there when called. What it does not buy is a different chip.
The discount tiers under list tell the same shortage story from the other side. At Google, the same A3 node drops to $50.32 an hour on Spot pricing and $38.86 on a three-year commitment. At AWS, P5 spot capacity barely exists at any price, because the on-demand pool itself is thin. Azure quotes a deep spot discount on paper, about $18.17 an hour for the same ND node, in the few regions where the capacity can actually be claimed. When every seller's interruptible tier is either expensive or empty, the market is not clearing at list. It is clearing above it, which is what the auction in the Nebius section below measured directly.
What do Amazon's own chips cost?
Amazon is the only hyperscaler that rents its own AI silicon to all comers, and it prices that silicon as the value tier of its own catalog. The lineup has three chip families, and each is sold on a different pricing model, which is itself part of the story.
Trainium1, the plain list. The first Trainium generation still carries a simple on-demand price. A single-chip trn1.2xlarge lists at $1.34 per hour, and the 16-chip trn1.32xlarge at $21.50, the same $1.34 per chip-hour. Through Capacity Blocks the same 16-chip instance runs $9.532 an hour, about $0.60 per chip-hour, less than half the on-demand rate.
Trainium2, the unlisted chip. The current training chip has no on-demand list price at all. AWS sells Trainium2 only through EC2 Capacity Blocks, prepaid reservations whose prices AWS states are updated regularly based on trends in supply and demand, with the next update scheduled for January 2027. The trn2.48xlarge, 16 chips in US East (Ohio), prices at $35.76 an hour, or $2.24 per chip-hour, about 37% of the $5.97 Capacity Block price of an H100 in the same program. A single-chip trn2.3xlarge exists in select regions at the same $2.24 per chip-hour. The missing on-demand price is itself a signal: the capacity is allocated to anchor tenants rather than sold openly, and the day Trainium carries a public on-demand price will mark a regime change, either supply outgrowing internal demand or AWS choosing to weaponize price in the open.
Trainium3, the allocated chip. The third generation is real: it appears in AWS's Neuron documentation at roughly 2,517 FP8 TFLOPS per chip, and Trn3 UltraServers scale to 144 chips in one system. It has no public per-hour price. Amazon's newest silicon is, for now, allocated to chosen customers rather than listed for rent.
Inferentia2, the quiet floor. The inference line keeps the simple on-demand list. The inf2.xlarge carries one Inferentia2 chip at $0.758 an hour. The inf2.48xlarge carries 12 chips at $12.98 an hour, about $1.08 per chip, the cheapest published per-chip rate on the ladder chart below.
The scale behind these prices is not a side project. The anchor deployment is Anthropic's Project Rainier in Indiana, an $11B campus where nearly half a million Trainium2 chips are already running Claude workloads, and AWS has said Claude is expected to run on more than a million Trainium2 chips. In April 2026 Anthropic committed to up to 5 gigawatts of Trainium capacity and more than $100B of AWS spending over the next decade. Amazon said in the first quarter of 2026 that its chips business, Trainium, Inferentia, and Graviton together, had passed a $20B annual revenue run rate and was growing at triple-digit rates. The quiet tier is one of the largest silicon businesses in the world, priced deliberately below the GPU market it competes with.
On the CPU side of the same strategy sits Graviton, Amazon's Arm server processor, priced about 15% below comparable x86 instances. It gets its own treatment in the CPU section below.
How does Google price the TPU ladder?
Google does the opposite of Amazon. Where AWS leaves its current Trainium unlisted, Google publishes a full TPU price ladder, per chip-hour, with every discount tier printed next to the headline rate.
| Google TPU, per chip-hour, US | On demand | 1-year commitment | 3-year commitment | Flex-start |
|---|---|---|---|---|
| TPU v5e | $1.20 | $0.84 | $0.54 | $0.60 |
| Trillium (v6e) | $2.70 | $1.89 | $1.22 | $1.35 |
| TPU v5p | $4.20 | $2.94 | $1.89 | $2.10 |
| Ironwood (v7) | $12.00 | $8.40 | $5.40 | $6.00 |
Read the ladder as a menu of trade-offs rather than a ranking. The same Trillium chip costs $2.70 on demand and $1.22 on a three-year commitment, a 55% cut for certainty. Ironwood costs 4.4 times Trillium's hourly rate and delivers about 2.5 times the BF16 throughput per chip with six times the memory, so its sticker is the highest on the ladder while its cost per unit of finished work is not.
The per-token arithmetic is where the ladder earns its keep. SemiAnalysis estimated Ironwood's inference total cost of ownership at $0.18 per million tokens in February 2026, against $0.31 for comparable B200 configurations, an advantage of about 42%. That gap, not the $12.00 sticker, is the context for Anthropic signing for up to a million TPUs in October 2025 and then for multiple gigawatts of next-generation TPU capacity in April 2026.
One widely repeated claim needs correcting here, because it shows how this market's numbers travel. Several 2026 write-ups describe an April 2026 Google Cloud release note cutting headline on-demand TPU pricing from $6.50 to $4.50 per TPU-hour, with preemptible capacity falling from $1.95 to $1.35. The figures are real, but the date is not: they come from July 2018, when Google cut the price of its original Cloud TPU. No equivalent headline cut appears anywhere in the 2026 ladder. Google's TPU prices this year were a ladder that held still while GPU prices climbed past it.
How much of the buildout is already custom silicon?
Ask how much of the AI buildout already runs on the buyers' own chips and you get three different answers, because there are three honest ways to measure it. The spread between the answers is where the argument about NVIDIA's position usually goes wrong.
| Custom silicon share, by ruler | 2024 | 2027 |
|---|---|---|
| Share of accelerator units shipped | about 30% | toward 50% |
| Share of TSMC CoWoS packaging allocation | about 18% | about 34% |
| Share of accelerator dollars spent | about 8% | about 20% |
Units flatter the challengers, because a cloud ASIC at a fraction of a merchant GPU's price counts the same as a Blackwell rack chip. Dollars are the ruler that touches revenue and margins, and on dollars the migration is real but mid-journey: custom silicon took about 8% of accelerator spending in 2024, about 12% in 2025, about 16% in 2026, and roughly 20% in 2027 on broker estimates, with forecasts reaching 35% to 40% of hyperscaler AI compute spend by 2028. NVIDIA still holds roughly 75% to 80% of the AI silicon wallet today. Bernstein's framing is the consensus one: that share slides toward 60% as the custom programs ramp, which is a demotion, not a displacement.
Meta wrote the end-state down in April 2026. Its Broadcom MTIA extension runs through 2029, an initial deployment above 1 gigawatt of custom silicon on a 2-nanometer process, and it was signed alongside roughly 6 gigawatts of AMD GPU capacity and millions of NVIDIA chips. That three-vendor structure is the destination for every large buyer: own silicon carries the base load of predictable work, merchant GPUs supply the flexible layer and the frontier, and no single vendor holds the whole wallet again.
The four buyers sit at different points on that road, and their positions explain their pricing behavior. Microsoft, with the least mature in-house program, remains the most dependent on NVIDIA and pays the most rent. Google, with the most complete program, has moved from customer to competitor and already exports TPU capacity to Anthropic. Two cautions belong on the record. Much of the headline capex in this buildout is land, shells, and power rather than silicon, so the silicon budget is smaller than the trillion-dollar totals suggest. And Google and Amazon book much of their custom silicon as internal cost flowing through design partners, which makes the true ASIC dollar figure run above what merchant revenue comparisons show. The direction survives both cautions. Only the speed is honestly debatable.
Why did CPU prices not move?
General-purpose compute is the control group in this experiment. A 4 vCPU instance with 16 GB of memory listed at about $0.19 an hour on AWS, about $0.19 on Google Cloud, and about $0.19 on Azure in April 2026 reference pricing, and comparisons published in August 2026 show the same numbers, matched to within fractions of a cent across the big three. CPU list prices are set in a mature market with deep spot and commitment discounts underneath them, and in a year when AI accelerators repriced by double digits twice, nobody moved them at all.
The deflation ran through performance instead. Every hyperscaler now fields its own Arm CPU: Graviton5 at AWS, Axion at Google, and Cobalt 100 at Azure. Each new generation held its hourly price roughly flat while lifting performance, so the cost of a unit of work fell without any price being cut. AWS prices Graviton at roughly 15% below comparable x86 instances, closer to 20% on memory-optimized families, and claims up to 40% better price-performance; independent benchmarks of the current Arm lines put the advantage at roughly 15% at AWS and Azure and slightly ahead at Google. A workload that migrated from x86 to Arm this year got cheaper the way compute used to get cheaper, silently, through the hardware.
The one CPU price move of the year ran the other way, and quietly. In June 2026 Oracle cut its Always Free Arm allowance in half, from 4 OCPUs and 24 GB of memory to 2 OCPUs and 12 GB, without an announcement. Even the free tier repriced.
Who is bringing back price discovery?
For most of this market, price is something a provider posts and a customer accepts. The exception of the year is Nebius, the neocloud that spent 2026 finding out what its capacity was actually worth, in public.
It started with an auction. Nebius disclosed its first capacity auction with its second-quarter 2026 results, and CEO Arkady Volozh described the outcome on the earnings call: "It was very successful and cleared at the highest price we have seen for the Blackwell generation of chips, 15% above the highest price we ever charged before. This gives us a strong signal on the value of this capacity in the market in real time." Chief revenue officer Marc Boroditsky added that the clearing price was 20% above the company's Blackwell pipeline pricing. The sequence matters. Nebius had already raised prices by hand, repeatedly, and demand had not blinked. The auction's message was that the manual increases had still been too small.
The posted prices followed. On October 1, 2026, Nebius raised on-demand rates by 17% to 21% across its NVIDIA line: the H100 from $3.85 to $4.50, the H200 from $4.50 to $5.40, the B200 from $7.15 to $8.50, and the B300 from $7.85 to $9.50. Its second-quarter shareholder letter had already reported pricing for older-generation GPUs running more than 30% above the first quarter. Management told investors it sold out of capacity as fast as new capacity came online, and that it could sell its 2027 capacity today at current deal terms but is holding some of it back for customers with immediate needs.
Then came the structural move. On October 8, 2026, Nebius switched its preemptible GPU capacity to dynamic spot pricing. The price is set per GPU type and per region from available capacity and demand, it can change as often as every 15 minutes, and it floats between a floor Nebius sets for each GPU type and a ceiling of the on-demand price minus $0.01. A customer either follows the spot price or sets a maximum price, and a machine whose cap is crossed stops. The prices are public without an account: seven days of history on the website and 30 days through the API.
Silicon Data sampled that API through the first nine days, October 8 to 16, and the newborn market behaved like a market. Nebius's H100 spot price sat at its $0.79 floor for most of the window, about two thirds below the $2.33 market spot median, and changed on only two of the eight measured days. Its H200, B200, and B300 prices moved on most days, mostly below market spot medians of $2.92, $4.01, and $4.95 respectively, and Nebius quoted below every other seller on most days for every GPU it lists. Set against the wider market, the contrast is sharp: more than half of all the spot price series Silicon Data tracks did not change once in the same nine days. Nebius is, so far, the only major seller whose prices actually float.
The industry's anchor used to run this experiment, and chose to stop. AWS built the original cloud spot market, then replaced its bid-driven pricing in November 2017 with a model that, in AWS's words, delivers prices that adjust gradually based on long-term trends in supply and demand. The spikes disappeared, and so did the information a floating price carries. AWS's accelerator capacity does reprice today, through Capacity Blocks, on a schedule measured in months; the current rates are next scheduled to update in January 2027. Nebius reprices in minutes and publishes the tape. In a shortage, that difference is worth money, and the auction measured how much: 15% above the best price the company's own sales process had ever achieved.
Is compute priced like credit now?
Read the sections above as one market and they mislead. There are two markets for the same silicon, separated by who the buyer is, and compute in 2026 is priced the way credit is priced: the rate you get is a judgment about you, not about the product.
At the top sit the anchor tenants, the Microsoft and Meta and Anthropic class. They are the 850-credit-score borrowers of this market. They sign multi-year take-or-pay contracts at the best effective rates in the system, they get silicon co-designed around their own models, and sellers compete for their signature. Their paper is so strong that it collateralizes the builder's own borrowing: in July 2026 Nebius closed a $775M secured facility priced at SOFR plus 250 basis points, backed by deployed GPUs and a contract with an investment-grade customer. A prime tenant's commitment is not just revenue to a neocloud. It is a credit-rating transplant that lowers the cost of the entire buildout behind it.
At the bottom sits the long tail: startups, mid-size labs, and anyone whose demand arrives in bursts. They get payday-loan terms. Spot capacity is instant and asks for no commitment, but it is capped at whatever the anchors have not taken, it floats up to a ceiling one cent below the on-demand price, and it is preemptible, which is repossession written into the product: when the seller needs the machine, the machine stops. The tail pays the highest rates in the market for the least certain product, exactly as subprime borrowers do, and for the same reason. It cannot credibly promise to be there in three years.
The tail is not a malfunction of the system. It is the equity tranche of it. Spot buyers absorb the volatility of the whole stack, the churn, the price swings, the preemption, so that the prime contracts above them stay pristine and bankable. That is why Nebius's little spot feed matters out of proportion to its size. It is the clearing price of the marginal buyer in the system, repriced every 15 minutes and published for anyone to read. Earnings calls report the top of the stack, where everything is always sold out. The feed reports the bottom. When this cycle turns, it will turn there first.
The price ladders in this article testify to something the demand narrative avoids: demand is finite. If buyers at the going price were unlimited, the rational seller would offer one product at one price and let a queue form. Instead every seller maintains a value ladder, flex-start at half of on-demand, four generations of TPU at four price points, an Inferentia line under everything, discounts for age. Perishability explains part of it, since an idle GPU-hour at 3 a.m. is worth nothing unsold. It does not explain a ladder this carefully cultivated. Sellers segment because demand is heterogeneous and limited, and they need a clearing price for every rung of willingness to pay.
The workloads explain why the segmentation holds. Compute demand is a pyramid. Its wide base, inference, fine-tuning, and ordinary training, is fungible: it runs on Trainium, TPU, Graviton, or AMD once an engineering team pays the adaptation cost, and the per-token numbers in the TPU section show the substitutes already winning that base on price. Its narrow apex, frontier training where time-to-model is the entire product and a slower run is a lost race, is where substitution genuinely fails. NVIDIA's pricing umbrella covers the apex by merit and the base by inertia, and the abstraction layers that let models move between chips get better every quarter. The umbrella does not collapse. It shrinks toward the apex.
Four kinds of demand. When an earnings call says demand is insatiable, it is blending four different things: contracted demand with signatures and upfront money, reserved demand that can still evaporate, announced demand measured in gigawatts nobody has financed, and vendor-financed demand that the seller's own balance sheet created. The standing test is to grade shortage by the price spread, never by the narrative. In 2024 the talk stayed hot while the H100 spot price halved, and that divergence was the fiction window. Scarcity also does financial work beyond price: take-or-pay terms and the 60% to 65% upfront payments Nebius enforces on term deals convert customers into financiers of the buildout itself.
The geography of the price tape tells the same story from another angle. If demand were truly worldwide, the spot price would follow the sun, rotating from Tokyo to Frankfurt to Virginia as FX prices do, and the valleys would fill behind each region's night. The compute feed instead runs on US hours and the US calendar. Regional list prices confirm the shape: Trillium lists at $2.70 in the US against $2.97 in Amsterdam and $3.24 in Tokyo, a premium for thin local supply, not evidence of local demand strength. What the worldwide-demand narrative resolves into, in prices, is American balance sheets buying at scale, a Chinese market sealed behind export controls at about 41% domestic vendor share, and sovereign buyers arriving as state-signed anchors rather than as market flow.
Where do prices go from here?
Start with what the trackers agree on, because it is the durable statement. The floor under AI compute moved up in late 2025 and is not returning to 2024 levels. The rate of increase moderated over the summer, JPMorgan's July decline was real, and utilization stayed above 70% while it happened. Providers were still raising posted prices in October. A 2027 budget built on the price collapse of 2024 is priced off the wrong year.
The constraint behind the floor is rotating. Through 2026 it has been memory: high-bandwidth memory supply caps how many accelerators can ship, whatever the demand. From 2027 it becomes power, as grid connections and on-site generation pace the data center buildout. After second-quarter 2026 earnings, all four hyperscalers reported 2027 capacity largely reserved and 2028 partly reserved, which means the marginal buyer next year is bidding for a small uncommitted remainder, in a market where one seller has just demonstrated what an auction does to the clearing price.
The sellers' own behavior adds a tell worth reading. Everyone in this market is pushing term: three-year commitments at Google, five-year anchors at the neoclouds, Capacity Blocks at AWS. A seller who expects tomorrow's price to be higher sells spot and keeps the upside. A seller racing to lock today's prices for years is publishing a private forecast, that the rates on offer now are worth insuring. The contracts also do balance-sheet work, since contracted backlog is the collateral the buildout borrows against, as the Nebius facility showed. That is the self-defeating loop inside every capacity cycle: the contracts signed to prove demand is permanent fund the capacity that eventually makes it ordinary.
On this article's own evidence, the cycle grades late-middle rather than final innings. Prices were still setting highs into the fall. The bottlenecks are physical and unresolved, memory this year and power next. No major default has yet revealed how much of the announced demand was phantom. And one quiet marker points the same way: sellers still hold a tranche of capacity back for premium short-term work, as Nebius does openly. At a true top, that tranche disappears into term contracts at any price a buyer will sign.
The strategic consequence runs straight into the companion piece to this article, The Compute Share Wars. Every dollar added to the rental price of NVIDIA compute improves the arithmetic of building your own chip. At a $2.24 Trainium2 chip-hour, and a published TPU ladder whose committed rates undercut the GPU medians, Amazon and Google are selling into a market whose reference price rose all year, and each program behind them, Meta's MTIA, Microsoft's Maia, and the efforts at ByteDance and OpenAI, gets easier to justify at these levels. Rising GPU prices are the custom silicon industry's best salesperson.
The price in one line. Compute stopped getting cheaper in 2026. The sellers with their own silicon now anchor the floor, CPU prices never moved at all, and the one seller running an auction found the market price 15% above its own best guess. The next thing to watch is whether anyone follows Nebius.
What happens to the chipmakers' margins?
Every price in this article is revenue to someone, and the reset eventually sends a bill to the sellers of silicon. In capacity cycles the bill arrives in a fixed order. Spot prices crack first, because they reprice in minutes. The contract index cracks second, as renewals replace scarcity-era paper. Backlog quality cracks third, in cancellations and weaker signatures. Reported margins crack last, because term contracts signed under scarcity keep reporting their economics for quarters after the market has turned. The 2026 tape has run that sequence in reverse, each gauge tightening in turn. The margin question is about the return trip, and about who holds a cushion when it starts.
NVIDIA's exposure is rent-sharing. Its gross margin is the whole stack's scarcity rent collected at a single toll booth, and three claimants now queue at that booth. The memory makers take the first cut, because HBM is the binding input and NVIDIA pays their price to ship at all. The buyers take the second cut, because every credible custom program in this article is a negotiation weapon at the next contract signing, whether it ships at scale or not. Mix takes the third cut quietly: rack-scale systems carry large amounts of other vendors' content through NVIDIA's income statement, diluting the margin percentage even while the dollars grow. None of this requires demand to fall. It only requires scarcity to stop being exclusively NVIDIA-shaped.
AMD's exposure is different. It cannot concede margin it is still trying to earn, and its seat at the frontier table has a price: the design wins that anchor its roadmap have come with warrants attached, tribute paid in equity rather than in price. Meanwhile its proven profit engine, EPYC server CPUs, is capped from above by the Graviton and Axion economics in the CPU section, where the hyperscalers' own Arm chips set the price of a unit of work. AMD has less margin to lose than NVIDIA. It also has less room to defend what it has.
Intel is the finished version of this story, arrived early through its own capital choices, and its third-quarter 2026 print shows both halves at once. The good half is real: data center revenue rose 78% as Xeon host processors rode inside other vendors' GPU racks, and gross margin recovered to 43.0% from a 32.7% trough. Intel participates in this cycle as a complement, priced by other people's scarcity. The other half arrived in the guidance. Intel guided fourth-quarter revenue to $15B to $16B against $17.07B of consensus, non-GAAP EPS to $0.25 against $0.43, and gross margin to 38%, with about 200 basis points of that guide being a hit from legacy Intel 7 server products. The stock fell about 25% on a good quarter, because the guide said the recovery has a ceiling. The full account is in the DalalBytes Intel investment report.
The deeper line in Intel's numbers is what its capital bought. The foundry rebuild consumed the war chest, $13.3B of segment losses in the peak year alone, and that foundry earned $411M from external customers in the third quarter. TSMC earned about $40.2B of foundry revenue in a single quarter of 2026. Margin, it turns out, is not a cushion in this industry. It is the currency the accelerator game is played with: it prepays for HBM allocation, wins packaging capacity, and finances the customers who need help signing. Intel is what a player looks like after the currency is spent. It can still ride the cycle. It can no longer set any of its terms.
What should you watch from here?
Calls about a cycle are only worth making if they can be graded. These are the tells that would move the readings in this article, stated in advance. None of them has happened yet.
THE WATCH LIST
1. Nebius's spot feed rolling over decisively, the marginal buyer repricing down for weeks rather than days.
2. New term deals pricing below spot, the inversion that says sellers value certainty over upside.
3. The spot tranche quietly vanishing from seller behavior, as every available hour gets pushed into term.
4. The first neocloud default, and a distressed GPU fleet coming to market.
5. Trainium receiving a public on-demand price.
6. The H100 contract index falling two consecutive quarters.
7. The first hyperscaler guiding capex growth below its own revenue growth.
8. NVIDIA's gross margin guide stepping down rather than being defended.
What is cheap here, and what only looks cheap?
Every price in this article eventually lands in someone's earnings, and the ten names that run this market do not stand in the same place when it does. The question the table below asks is not which multiple is lowest. It is where each company's earnings denominator sits in the cycle, because a multiple is only as honest as the earnings underneath it.
| Company | Price | Fwd P/E | Fwd rev growth | Gross margin | Earnings base | Score |
|---|---|---|---|---|---|---|
| Broadcom (AVGO) | $326.08 | 18.7x | +76% | 74.2% | Still ramping | 76/100 |
| Microsoft (MSFT) | $521.91 | 25.1x | +20% | 67.8% | Depressed by capex | 74/100 |
| Alphabet (GOOGL) | $279.84 | 20.3x | +27% | 59.7% | Depressed by capex | 74/100 |
| Amazon (AMZN) | $240.35 | 24.3x | +16% | 53.2% | Depressed by capex | 74/100 |
| Meta (META) | $614.99 | 18.7x | +23% | 79.3% | Depressed by capex | 80/100 |
| Nebius (NBIS) | $154.14 | n/a | +429% | 77.1% | Still ramping | 70/100 |
| Arm (ARM) | $205.97 | 86.2x | +26% | 97.2% | Still ramping | 54/100 |
| AMD (AMD) | $504.80 | 45.0x | +64% | 56.0% | At cycle peak | 72/100 |
| NVIDIA (NVDA) | $204.64 | 16.9x | +78% | 75.0% | At cycle peak | 76/100 |
| Intel (INTC) | $65.76 | 46.9x | +9% | 43.0% | Impaired | 47/100 |
Prices are the Friday, October 30, 2026 close. Forward P/E and forward revenue growth are analyst consensus estimates from StockAnalysis, snapshot November 2, 2026; the growth column is consensus revenue for the next twelve months measured against the trailing twelve months, implied by each stock's price-to-sales and forward price-to-sales ratios on that snapshot. Gross margin is each company's latest reported quarter. Scores are DalalBytes five-pillar research scores, shown where a house report exists.
Broadcom screens cheapest on this table, and the reason is the denominator. At 18.7 times forward earnings it carries a hyperscaler's multiple while its earnings are still ramping into a contracted guide, a fiscal 2027 AI revenue guide of $115B from a management team with a record of guiding low, against a trailing base that has not yet printed those quarters. The multiple is low because the E has not arrived yet, not because the market doubts it will. The hyperscalers sit one rung down. Microsoft at 25.1 times, Amazon at 24.3, Alphabet at 20.3, and Meta at 18.7 are all being priced through the capex trough of their cash cycles, at peak spending intensity and trough free cash flow. October did the repricing: Alphabet fell 18.7% in the month and Meta 15.2%, on prints whose revenues beat. Their earnings are depressed by the buildout, not impaired by it, and the cash engines underneath never stopped compounding.
Nebius does not belong on a P/E table at all, and the n/a in its row is the honest entry. It is priced as a toll road, and what matters is the quality of the paper on the road: the anchor contracts with investment-grade tenants, the 60% to 65% upfront payments on term deals, the secured borrowing those contracts collateralize. Its consensus revenue more than quintuples over the next twelve months on the same snapshot, which is another way of saying the trailing numbers describe a company that no longer exists. Arm is the opposite shape. Its royalty stream carries a 97.2% gross margin, the most defensible margin structure in this stack, because Arm sells no silicon and rents no capacity, and its earnings base is still ramping as AI designs ship. The business is right. The price is wrong: 86.2 times forward earnings asks the buyer to prepay a decade of that ramp today.
AMD and NVIDIA are the inversion, and NVIDIA's row is the trap in its purest form. At 16.9 times forward earnings NVIDIA carries the lowest multiple on this table, and it means the least, because the earnings underneath it are printed at a 75.0% gross margin at the top of a pricing cycle this article has just spent eleven sections describing. AMD's 45.0 times looks expensive and is, but its denominator sits at the same altitude, a 56.0% gross margin earned while it is still buying its seat at the frontier table. This is the classic cyclical trap. The dip is in the price. The earnings have further to fall, through the rent-sharing sequence the margin section laid out, and a low multiple on peak earnings is how cyclicals look cheapest at exactly the wrong moment. Intel is the finished version of the same story, arrived early. Its margin currency was spent on the foundry rebuild, its latest quarter printed a 43.0% gross margin against a 38% guide for the next, and at 46.9 times forward earnings the market is paying a growth multiple for recovery hope. Intel is not cheap on any denominator that exists yet.
The ranking that falls out of the table is a ranking of denominators, not of businesses. Broadcom first, because its earnings are still arriving. The hyperscalers next, because October repriced them on trough cash flow at peak capex while their engines kept compounding. Nebius beside them on different evidence, the quality of its contract paper rather than any multiple. Arm after that, a right business waiting for a wrong price to correct. AMD and NVIDIA last among the operating names, not because their franchises are weak but because their multiples are quoted on peak earnings, and Intel below them all, priced on a recovery its spent margin can no longer finance. Cheap, in this market, is a statement about where the E sits. The table is the whole argument.
Frequently asked questions
Did GPU rental prices go up or down in 2026?
Up. After two years of decline, the H100 one-year contract index bottomed at $1.70 per GPU-hour in October 2025 and rose about 40% by March 2026. Neocloud indexes and on-demand medians climbed through October 2026, and Lambda, Runpod, Nebius, and AWS Capacity Blocks all raised list prices into the fall.
Is the AI compute shortage real or manufactured?
Both, in layers. The shortage is real at the bottleneck, where HBM supply this year and power from 2027 cap how much capacity can exist. It is managed at the margin, through allocation to anchor tenants and tiered pricing that rations the rest. Grade it by the price spread between contract and spot, never by the narrative.
Why do hyperscalers charge so much more than neoclouds for the same GPU?
The GPU is bundled with everything around it. Hyperscaler prices carry enterprise agreements, global regions, networking, storage, support, and capacity that is actually there when called. The same eight-GPU H100 node lists at $55.04 an hour on AWS and $98.32 on Azure, against $49.24 at CoreWeave and a $2.82 per GPU neocloud index.
Disclosures
Not financial advice. This article is published for educational purposes only. Nothing here is a recommendation to buy, sell, or hold any security, and nothing here is financial advice. Investing involves risk, including the possible loss of principal.
Positions. The author holds long positions in Amazon, Alphabet, AMD, Meta, and Nebius, all discussed in this article. The author holds no position in NVIDIA, Microsoft, Intel, Broadcom, Arm, or Marvell.
Accuracy. Compute price figures come from the trackers and providers named beside them: the SemiAnalysis one-year H100 rental index as reported; JPMorgan's July 2026 on-demand reading as reported; Silicon Data's neocloud rental indexes and its October 8 to 16, 2026 study of Nebius spot pricing; GetDeploying's provider medians; Ornn's transaction index; and provider price lists checked in October 2026, including AWS on-demand and Capacity Blocks pricing, Google Cloud TPU pricing, Azure and CoreWeave list prices, and the Lambda, Runpod, and Nebius price pages. Nebius auction and spot pricing details are from the company's second-quarter 2026 earnings call, its shareholder letter, and its published documentation, and the terms of its July 2026 secured facility are from company reporting of that closing. Amazon chip figures are from AWS pricing pages and company statements. Custom silicon share figures are broker and research-firm estimates as reported, including Bernstein's hyperscaler wallet framing and TrendForce shipment and packaging allocation data, and are estimates, not census counts. Intel figures are from Intel's third-quarter 2026 earnings release and guidance and from the DalalBytes Intel investment report. The credit tiering, cycle grading, and margin sequence readings are the author's analysis of the price data presented, not reported fact. The July 2018 Cloud TPU price cut is from Google Cloud's own release notes of that date. Prices in this market change often and by region and commitment; figures here are dated snapshots, not quotes. Errors can occur and facts change; verify independently before acting on anything you read.
No advisory relationship. Reading this site does not create an advisor-client relationship of any kind. See the full disclosures page.