The same H100, six different prices, and the most expensive one is the default. A tour of everything wrong with hyperscaler GPU billing.
Hot Takes
Short, sharp, arguableThey mock the tokens in public and rent the GPUs in private. Follow the purchase orders, not the press releases.
Deep Dives
Long-form breakdownsReasoning models and agents turned inference from a bursty call into a continuous, elastic load. Why the workload that's growing fastest is the one decentralized networks are best suited to serve.
From FLOPs to final invoice: what a training run actually costs when you count idle time, failures and egress — not just the brochure rate.
Subsidized supply is not demand. Why most decentralized compute networks would die in a month without emissions — and what a real one looks like.
Every quote is thousands of simulated worlds. Why actuarial compute is peaky, wasteful, and shaped exactly like the load decentralized GPU markets were built for.
Chips ship in months, interconnection queues take years. Why the fastest new AI capacity comes from stranded power and retired industrial sites.
Bulls and bears both assume today's GPUs stop being useful. Depreciation and obsolescence are two clocks running at wildly different speeds — and only one of them can be a bubble.
Meta, Microsoft, Amazon and Google have signed for ~9.8 GW of advanced nuclear. Almost none of it exists yet — and that gap is the real compute story of 2026.
Decode runs at about 1 FLOP per byte against a balance point near 200. That ratio, plus HBM's three-times wafer cost per bit, explains why $/hr and $/token disagree.
A single GB200 NVL72 rack holds 72 Blackwell GPUs and draws ~120 kW — past what air cooling can shed. Why the real limit on AI compute in 2026 is the building, not the chip.
We found the bottleneck in GPUs, HBM, power, the building, the package and the water. The one inside all of them is quieter: accelerators are billed at 100% and used at a fraction of it. That gap is the utilization tax, and it is where the real, bookable supply is hiding.
US data centers consumed 66 billion liters of water on site in 2023 and about 800 billion litres generating their electricity — a 12× multiplier that almost never appears in a sustainability deck. Once you price that, "closed loop or evaporative" stops being a values question and becomes arithmetic with a break-even you can calculate.
A 5.5x-reticle interposer yields single-digit packages per 300mm wafer, and the insulating film holding the substrate together is ~95% one supplier. Why packaging — not wafers — rations accelerators.
Intel's 2023 test chip claims 10x interconnect density; TSMC's June 2026 supply-chain plan tests 0.8mm glass cores at 85x110mm; Absolics builds on a NIST-backed $75M award. The next packaging wall may be glass — a stage-by-stage read of who has published what.
Batteries catch the load in milliseconds, generators must accept it within 10 seconds (NFPA 110 Type 10), fuel decides the rest. Uptime's 2026 data: 57% of major outages cost over $100k — and the backup chain itself fails more often than the grid.
1.8 TB/s inside the rack, 100 GB/s leaving it: a 9x cliff every training job pays for. Radix math (N²/2 endpoints), two links per GPU, and the five fabric questions a GPU cloud should be able to answer.
Meta's 405B run lost 419 jobs in 54 days — one every 3.1 hours. Mixed-precision Adam carries 16 bytes per parameter (6.5 TB per checkpoint), and Young/Daly puts the optimal save interval at ~13.6 minutes with a 7.3% floor on wall clock.
An agent's bill has four segments: fresh input, cached input at ~0.1x, visible output, and reasoning tokens billed as output but never shown. Anthropic measured 4x/15x multipliers; our arithmetic says raw token counts overstate the input bill by ~5x when caching works.
Server DDR5 contract prices up 93–98% in one quarter of 2026, spot DDR5 die up 307% in four months, and Micron walked away from its own 29-year-old consumer brand. The AI memory shortage is displacement, not addition — and everyone downstream pays.
One exabyte rents for $23M a month at S3 Standard rack rate, freezes for ~$1M in Deep Archive, costs ~$7M at Backblaze — and the media behind it amortizes to ~$0.3M off Seagate's record quarter. A 20–70× spread, plus a $90M toll to read it back out.
One 256 GB module adds 32 GB/s and 2–2.5× local DRAM latency — about 5% of a Xeon 6 socket's bandwidth. Eight fill the lanes with 2 TB and ~250 GB/s. The KV-cache arithmetic that decides which side of the line your workload sits on.
896 W of pluggables against an ASIC that draws less, and 1,792 W at the next generation. CPO cuts 60–70% of that — about 1% of a campus. The density argument, the lasers that stay outside, and the commodity you stop being able to shop for.
1 MW at 54 V is 18,519 A and 332 kg of copper per metre; at 800 V it is 1,250 A and 22 kg. But the busbar only burns 0.06% of a gigawatt — so the 3–5% efficiency claims are about deleted conversion stages, and Google and NVIDIA quote different numbers because one is a retrofit.
720 Megapacks, 2.8–3.3 GWh, behind the meter at one AI campus — against a 25 MWh ride-through for a 100 MW hall. The battery stopped being a UPS and became a load-shifting generator, and an instrument for buying grid access.
Sub-1-MW buyers paid $293/kW-month in Digital Realty's Q2 2026 Americas signings while above-1-MW paid $156 — an 88% premium for being small. Renewals repriced +25.4% in one quarter while headline asking rents moved 1.5%: the quoted rate is not the number that decides your bill.
Alibaba Cloud's CUBE 5.0 delivers a large AI data center in 100 days (30 factory + 50 install + 20 commissioning) against a 12–18-month US norm — but the window starts at a powered shell and ends before IT arrives, while grid interconnection averages 4+ years. Modularization compresses the build, not the queue; our arithmetic says the speed benefit ≈ the cost benefit.
One GB200 NVL72 rack — 72 Blackwell GPUs in a 130 TB/s NVLink domain — rents at three official prices: CoreWeave $10.50, AWS $10.582, Azure on-demand $27.04 per GPU-hour. Azure's own 3-year commitment collapses the anchor to $11.90; the ~$10.4M per-rack-year spread is one megawatt of construction. List is an anchor; commitments are the market.
PJM's 2027/28 capacity auction cleared at the FERC cap of $333.44/MW-day — 11.5× the 2024/25 price — and still came up 6,623 MW short. Capacity bills the megawatt you reserve, not the MWh you burn: $15.44/MWh at 90% load factor, $46.31 at 30%. Ohio's 85% floor for 12 years and Virginia's $1.5M/MW collateral turn that line into a balance-sheet commitment.
GE Vernova ended Q3 2026 with 122 GW of gas equipment backlog plus slot reservations — 6.1 years of output at its 20 GW run rate, which lands in 2032, the delivery year it sold slots for. Siemens Energy closed fiscal 2026 at 104 GW, 39 of them reservations. Data centers are 10–20% of buyers and 100% of the marginal price: the scarce product is a date, not a machine.
NVIDIA's 2026-03-02 investment of $2 billion each in Lumentum and Coherent equals roughly 4.4× Lumentum's entire fiscal-2026 capex of $451.3 million — and Lumentum's revenue ran $480.7M → $1.01B → $1.26B in twelve months while net PP&E grew 60%. The constraint on 800G/1.6T optics is wafer time in compound-semiconductor fabs, new capacity lands after the 1.6T ramp, and China refines ~70% of the world's indium. The $4B is option premium on queue position.
Dell's AI server revenue doubled to $16.4B with a 20.9% gross margin — and free cash flow fell 47%. Supermicro shipped $19.9B at a 12.7% gross margin and used $434M of operating cash, with receivables up $3.56B and inventory up $2.48B. The margin is rent on allocation; the constraint is working capital, and Dell's $95B backlog is 5.8 quarters of demand priced before the memory bill arrives.
Micron ended fiscal 2026 with 26 strategic customer agreements — take-or-pay deals covering 35%+ of revenue through 2030 — and ~$150 billion of RPO counted at minimum prices, backed by $32 billion of customer deposits ($12.7 billion on its balance sheet). SK hynix signed 14 LTAs covering 50%+ of 2027–2028 capacity with demand at 2× supply. The spot price is becoming the residual, and the floor is struck at record-cycle margins.
The $197B backlog splits into $104B equipment (+92% YoY) and $93B services (+14%) — 47% of the order book is annuity, not turbines. Services backlog grows twice as fast as services revenue converts, customer down payments drove $5.3B of the quarter's $5.4B free cash flow, and every one of the 32 turbines shipped is a future service contract being commissioned.
Bloom Energy's Q3 2026 was $1,235.8M of revenue — 86.9% product, 5.5% service, at a 33.8% gross margin. GE Vernova books 47% of its backlog as a services annuity; Bloom books a machine. What buyers want is a date: 1 GW at $40B over four years is $27.4M a day of asset-time against a ~$280M-a-year power premium.
Pricing Watch
Real GPU market rates, trackedWe pulled on-demand rates from five hyperscalers and three decentralized marketplaces. The spread will make you angry.
The old rule was "always rent". The new rule needs a spreadsheet. Depreciation curves, utilization thresholds and the break-even points.
Used H100 prices span $6,000 to $25,750 across sources. Ask prices, executed prices and self-reported residuals get reported interchangeably — here's how to read the numbers honestly.
Same accelerator-hour, two prices. The break-even utilization nobody computes, and the hybrid pattern that quietly saves 25-45%.
The AI buildout was financed, not funded. Underneath the debt sits a collateral value — an accelerator's residual worth — that has never had a public price.
Landscape
Who's who, reviewed honestlyA 1 GW campus needs 4,000+ construction workers at peak and runs with roughly 300 people once it opens. Why the published ratios disagree by 6×, and why one month of slip costs about one month of rent.
Export law governs chips crossing borders, not compute accessed remotely. That mismatch split the market into compliant capacity and shadow capacity — and both are now priced in.
Every project publishes "registered GPUs". Almost none publish paid hours. A unsentimental tour of who is real and who is a spreadsheet.
Nineteen startups, one metaphor, zero Airbnbs. What the marketplace metaphor hides — and which teams escaped it.
180+ national AI programmes, yet 90% of frontier compute sits in two countries. What "sovereign compute" actually buys — and where a decentralised network can sell into it.
LBNL counts power plants asking to sell. The load queue is separate and far bigger: ERCOT alone reports 438 GW of large-load requests, 89% of them data centres.
Power transformers: 128-week lead times, GSUs 144, prices +77% since 2019. The equipment queue sits after the grid queue — and it is the line item missing from every capacity model.
On-Chain Gaming
Games, economies — and their compute billsThe autopsies are done: paying players to play selects for players who hate games. What the survivors rebuilt on the ashes.
Out of hundreds of "AAA web3 games" promised since 2021, five are live, patched weekly and retain players. A field report.
Every studio wants ChatGPT-grade villagers. Nobody budgeted for them. The hidden line item that connects game servers to decentralized compute.
Five pairs a day. One photograph is real, one was generated. We built it because half the arguments in this forum come down to "looks real to me" — now there is a scoreboard.
Homelab
Your own machine as the datacenterA used RTX 3090 is about $820. Whether it beats renting depends on three numbers: your utilization, your electricity price and your exit. The full arithmetic, line by line.
24 GB of VRAM for around $800, running quantized models that didn't exist two years ago. Why Ampere keeps winning the home-lab argument in 2026.
Storage-first or VM-first? The honest fork in the road for beginners — and the classic combo that runs both.
The GGUF file on your disk is roughly two-thirds of what actually runs. Why the KV cache — not the weights — decides which model fits your card, with the arithmetic for 14B, 32B and 70B.
Residential bandwidth, dynamic IPs and electricity bills vs marketplace rates. The realistic number is smaller than the marketing — and the real ROI is elsewhere.
2.5G was standardised in 2016 to run over the Cat5e you already own. 10G is where cable becomes the spec — and where the drive, not the network, becomes your ceiling.
1 W held for a year is 8.76 kWh. 80 Plus only certifies 20%, 50% and 100% of rated output — so the badge says nothing about the light-load region a 24/7 server actually lives in.
A UPS is sold in VA but your rack draws watts — and APC's own docs put the industry rule at "approximately 60% of the VA rating." The two-limit arithmetic, the runtime curve, and the checklist for sizing backup power honestly.
Micron rates the same 30.72 TB drive at 56,000 TB or 14,000 TB depending on the write pattern — a 4× spread on one datasheet. The DWPD formula, why the client class assumes 8 hours a day at 40°C, and the one number to measure before buying used flash.
NEC 210.20(A) sizes the breaker at the noncontinuous load plus 125% of the continuous load — and a 24/7 rack is a continuous load by definition. So a 20 A / 120 V circuit is 1,920 W continuous, not 2,400 W: the load tally, the receptacle math, the wire gauge, and the 240 V option that doubles your watts on the same copper.
1,000 W of compute is 3,412 Btu/h of heat and needs about 316 CFM at a 10 °F rise — roughly six 120 mm fans at free air. Fans only set the temperature difference, and the fan laws charge 8× the power for double the airflow. The equation, the ASHRAE envelope, and why 30 m³ of room air warms a kelvin every 36 seconds.
A rack PDU has three amp limits and most people check one. The branch allows 1,920 W continuous, but the C13 connector is rated 10 A by IEC 60320 (12 A on Vertiv's own UL tables) — about 1,440 W per outlet at 120 V — and C19 steps to 16 A. The failure mode is a warm plug and intermittent reboots months later, never a breaker trip.
A line-interactive UPS pauses its output for 2–4 ms on transfer; the Intel ATX12V spec requires a power supply to hold regulated output for 17 ms at full load, and ATX 3.1 still mandates 12 ms. The bridge your PSU was born with covers the UPS's gap three to six times over — which is why most home racks need AVR and correct sizing, not double conversion.
AM5 exposes 24 usable PCIe 5.0 lanes, Xeon 6 gives 96, EPYC 9005 gives 128 per socket, Threadripper PRO with WRX90 gives 144. Four GPUs at x16 need 64 of them before storage. And x8 versus x16 barely moves tokens per second — lanes buy load time and headroom, not throughput.
Guides
Hands-on: get set up, step by stepChain ID 97, RPC URL, symbol, faucet — plus importing the tOMC test token. Every value read from the chain on 2026-10-09, not copied from a screenshot.
Two of the three tokens are worth nothing by design. The airdrop rules — 20M OMC, 1M slots, 10 entries per wallet, Nov 1 — in plain numbers.
Any CUDA or ROCm GPU works. Staking tiers, heartbeats, what gets slashed — and why the testnet buys reputation, not income.
About this forum
The Decentralized Compute Forum (DCF) is an editorial publication about GPU markets, DePIN networks and the economics of AI compute. We write it, we argue in it, we publish our sources — and we change our minds in public when the data does.
No paywall, no sponsored posts. Comments run on GitHub Discussions — github.com/CBTdao.
Full disclosure: DCF is operated by the Omniverse Compute (OMC) team — a decentralized GPU network project on BNB Chain, currently in public testnet. Coverage of OMC and its competitors follows the same rules as everything else here: dated numbers, primary sources, public corrections. The privacy policy, terms and about page carry the details.