GPU Clouds in 2026: H100s at $2.59 an Hour, $104B of CoreWeave Backlog on $35B of Debt, and a Six-Year Bet on Silicon

I rent GPUs the way most people reading this do: in small quantities, on demand, from whoever has them this week. PandaStack, the Firecracker microVM cloud I build and run (my company, so weigh what follows accordingly), is CPU-heavy, and its GPU footprint is a line item I negotiate rather than a fleet I own. That puts me on the buy side of a market that in three years has gone from allocation by relationship to a transacted price index, and from venture equity to investment-grade bonds secured on graphics cards. I spent the week in the 10-Qs, the 6-Ks, the loan announcements and the price pages. This is what the GPU cloud looks like in October 2026: what an hour costs, who owes what to whom, how long everyone claims the chips will last, and at what utilisation owning beats renting.

The price did not go to zero, and then it went up

The headline everyone remembers is the fall. In spring 2023 an eight-GPU AWS P5 node listed above $60 an hour, about $7.50 per H100-hour, per emma.ms's provider survey. Coloprice's compilation of Silicon Data's index puts mid-2023 at roughly $8, 2024 at $3 to $5, the first half of 2025 at $2 to $3, and the October 2025 trough at $1.70, when the market briefly believed Blackwell would make Hopper worthless. SemiAnalysis's December 2023 model had shown that an operator paying 13 percent interest needed about $1.53 an hour to cover an H100 server's capital and hosting, which is why $1.70 was a floor.

Less reported is the rebound: $2.00 in December 2025, $2.20 four weeks later, SemiAnalysis's one-year contract index at $2.35 in March 2026, Silicon Data around $2.53 by mid-year. Ornn's OCPI-H100, a volume-weighted index of transacted neocloud H100 SXM rentals rather than rate cards, printed $2.59 per GPU-hour on 4 October 2026, with a three-month range of $2.46 to $3.17. The operators say the same. CoreWeave told its Q2 2026 call that prior-generation pricing "is at or above where it was years ago", that it raised list prices about 25 percent in July, and that it had signed an A100 contract, a 2020 architecture, running into 2029. Nebius's Q2 letter reports deals above $20 million of annual revenue per megawatt and "more than 30% higher pricing on older-generation GPUs versus Q1".

The spread between the transacted index and the rate cards is the second thing to notice. Here is what the public price pages said on 5 October 2026, per GPU per hour, before any negotiated discount.

Provider, page fetched 5 Oct 2026	H100 SXM $/GPU-h	H200	B200	Notes
CoreWeave, HGX 8-GPU	$6.16 on-demand, $2.46 spot	$6.31	$8.60	GB200 NVL72 $10.50 per GPU
Lambda, 8-GPU	$3.99 ($4.29 single)	not listed	$6.69	reserved via sales
Nebius	$3.85, rising to $4.50 from 1 Oct	$4.50 to $5.40	$7.15 to $8.50	preemptible from $0.79
Together	$3.99, $3.19 at 91 to 180 days	$5.99	$8.19	GB200 "contact"
RunPod	$2.69 community, $3.49 secure	$3.59 / $4.59	$5.98 / $6.79	community = third-party hosts
Vast.ai, public offers API, 1x on-demand	low $1.98, median $3.79, 6 offers	n/a	n/a	query I ran on 5 Oct
OCPI-H100 transacted index	$2.59 on 4 Oct, 30-day -10.4%	n/a	n/a	volume-weighted neocloud trades

SF Compute's market page redirects to a login and Shadeform's directory is client-rendered, so both are absent rather than guessed. The pattern is clear anyway: on-demand list is $3.85 to $6.16 at the integrated neoclouds, the marketplace floor is around $2, the transacted average is $2.59, and the newest silicon carries roughly a 2x premium (B200 at $6 to $8.60, GB200 at $10.50) because buyers pay for tokens per dollar, not hours. Nebius's 1 October increase and CoreWeave's July rise are the first across-the-board list price rises since 2023.

What the neoclouds filed

"Neocloud" covers everything from CoreWeave, with $35 billion of debt and bond ratings, to companies that have never disclosed revenue. The table is what I could verify from filings and primary announcements.

Company	Period	Revenue and result	Contracted	Debt and capital	Filing
CoreWeave (CRWV, IPO Mar 2025)	Q2 2026	$2,575M (+112%), net loss $626M, interest $640M	RPO $103.7B; 1.5 GW active, 4.2 GW contracted	debt $35.1B, cash $5.5B, FY26 capex guide $35B to $39B	10-Q
Nebius (NBIS)	Q2 2026	$582.3M (+454%), adj. EBITDA $236.2M, net loss $190.4M	>$40B commitments; 5 GW contracted power targeted by year-end	debt and converts $8.55B plus $5.75B Aug converts; cash $8.0B; Q2 capex $5.66B	6-K
Oracle (OCI)	Q1 FY27 to 31 Aug 2026	$19.3B, OCI $7.4B (+121%), FCF negative $5B	RPO $664B, 13% within 12 months	borrowings $125.3B, quarterly capex $28.5B	8-K, 10-Q
Lambda (private)	2026	reported >$1.5B expected; nothing audited public	two investment-grade offtakers behind the October loan	$926M TLB at SOFR+3.00 (Aug); $1.008B at 6.78% fixed, Baa1 (Oct); IPO reported "as soon as 2027"	Lambda, TFN
Crusoe (private)	Sept 2026	not disclosed	$140B+ TCV, 6 GW+ contracted, 1 GW live	$3.9B Series F at $30.9B	Crusoe
Together AI (private)	July 2026	not disclosed	>500 MW investor-capitalised commitments	$800M Series C	Together
Fluidstack (private)	2026	not disclosed	Anthropic's $50B Texas and New York build	reported talks at $18B valuation in April	Anthropic, TechCrunch

Two numbers deserve a second look. CoreWeave's backlog went from $60.7 billion at the end of 2025, per the 10-K, to $103.7 billion six months later, against 2026 revenue guidance of $12.4 to $13.2 billion: about eight years of current revenue is contracted. Oracle's $664 billion RPO against $19.3 billion of quarterly revenue is more extreme, and only 13 percent converts within twelve months. Both spend ahead of conversion: Oracle's free cash flow was negative $5 billion on $28.5 billion of quarterly capex, and CoreWeave's 2026 capex guidance is three times its revenue.

Concentration, and the circle the money runs in

The counterparties behind those backlogs are few. CoreWeave's 10-K says Microsoft was about 67 percent of 2025 revenue; the Q2 10-Q shows Customer A at 36 percent and Customer B at 26 percent, with A down from 71 percent a year earlier, so diversification is real but the top two are still 62 percent. The 10-K names an OpenAI agreement of up to $11.9 billion through October 2030, since expanded to about $22.4 billion per CoreWeave's newsroom, and a Meta order of up to $14.2 billion through December 2031. Microsoft anchors the others too: Nebius's September 2025 6-K discloses a five-year agreement for dedicated capacity at Vineland, New Jersey, worth about $17.4 billion through 2031 and up to $19.4 billion if Microsoft takes more. Introl's tally, which I could not confirm from Microsoft itself, adds roughly $23 billion to Nscale for about 200,000 GB300s, $9.7 billion to IREN and a "multi-billion" Lambda deal, for more than $60 billion; IREN's investor news confirms it delivered its "Horizon 1" site to Microsoft in August 2026.

Then the circle. NVIDIA's September 2025 letter of intent is to invest up to $100 billion in OpenAI, released gigawatt by gigawatt as 10 GW of NVIDIA systems deploy, the first on Vera Rubin in the second half of 2026. Bloomberg's reconstruction, quoted by Noah Smith, adds NVIDIA's purchase of $6.3 billion of unsold CoreWeave capacity and OpenAI's $300 billion Oracle contract. NVIDIA's 10-Q to 26 July 2026 shows the balance sheet version: non-marketable equity up from $22.3 billion in January to $47.9 billion; $366 billion of commitments, including $29 billion of cloud service agreements, NVIDIA renting back its own chips; and August guarantees capped at $105 billion behind SB Energy's leases of about 4.25 GW to an OpenAI affiliate. The vendor is the customer's largest investor, a buyer of the customer's customers' capacity, and guarantor of the customer's landlord. Vendor financing built the railways, so none of this is novel, but the demand the neoclouds borrow against is partly demand NVIDIA pays to create, and a lender should know how much of an offtaker's cash came from the chip supplier.

One deal in this web failed, instructively. CoreWeave's $9 billion all-stock purchase of Core Scientific and its 1.3 GW of contracted power was voted down on 30 October 2025 because the 0.1235 exchange ratio had no downside collar. Power owners preferred being CoreWeave's landlord to its shareholder, which says where they think the scarce asset is; I made the same point about substations in AI Power in 2026.

The debt stack, from 15 percent to 5.9

The most interesting engineering in the sector is a loan secured on GPUs going from private-credit curiosity to investment-grade asset class in three years. CoreWeave's first facility, $2.3 billion led by Magnetar and Blackstone in August 2023, was secured on H100 servers and their contracts. The Q2 10-Q's debt table shows what that vintage cost: DDTL 1.0, maturing March 2028, carries a 15 percent effective rate; DDTL 2.0 is at 11 percent; the senior notes of 2025 and 2026 carry 9.00, 9.625 and 9.75 percent coupons.

Then the curve bent. The DDTL 4.0 facility, $8.5 billion closed on 31 March 2026 with Blackstone as anchor, was rated A3 by Moody's and A (low) by DBRS and priced at SOFR plus 2.25 percent floating or about 5.9 percent fixed, maturing 2032. The 10-Q describes the structure: a bankruptcy-remote subsidiary, non-recourse to the parent, secured on its equipment and contracted cash flows, advancing against "the depreciable cost of computing equipment, projected debt service coverage and project-level conditions", with 95 percent of floating exposure hedged. The $3.1 billion DDTL 5.0 in May was the first publicly syndicated, at SOFR plus 4.50; the $2.6 billion DDTL 5.5 in August, at SOFR plus 5.50, was the first to accept shorter-dated contracts, which CoreWeave's CDO said "allows us to target a wider variety of customers, including global enterprises that typically favor shorter-term agreements". The gap is how the market prices a five-year Meta contract against a two-year enterprise one: roughly 325 basis points.

Lambda followed in miniature: a $926 million term loan B at SOFR plus 3.00, Baa2, closed 27 August 2026 and secured on "GPU servers and related infrastructure funded through the transaction and the cash flows those assets generate"; then a $1.008 billion fixed-rate facility at 6.78 percent on 1 October, rated Baa1 and A (low), amortising to 2033, sold to insurers. Nebius's first secured deal in July was $775 million at SOFR plus 2.50, followed by $5.75 billion of convertibles in August.

The cash cost is on CoreWeave's income statement: $640 million of interest in Q2 2026 against $2.58 billion of revenue, guided to $860 to $940 million in Q3, with 2025 interest of $1.23 billion already more than triple 2024. Management says the weighted average cost of debt fell almost 300 basis points in a year, and that a typical five-year contract repays the asset-level debt within its term, leaving an unencumbered cluster to re-rent. That claim is the business model, and it rests on two things a lender cannot verify from the collateral: that the offtaker pays for five years, and that the cluster is worth renting in year six.

Six years or three

Which brings us to the argument Michael Burry started in November 2025. His claim, as summarised by Level-Headed Investing and Dave Friedman since I cannot link his posts on X, is that the hyperscalers depreciate GPUs over five to six years when the economic life is two to three, understating depreciation by about $176 billion across 2026 to 2028. The useful lives themselves are not in dispute; they are in the filings.

Company	Server useful life on the books	Change and stated effect	Filing
Microsoft	two to six years	none disclosed in FY2026	10-K FY2026
Alphabet	"generally over a period of six years"	extended to six in 2023	10-K 2025
Meta	5.5 years from 1 Jan 2025	2025 depreciation down $2.92B, net income up $2.59B	10-K 2025
Amazon	a subset cut from six to five years from 1 Jan 2025, citing "the increased pace of technology development, particularly in the area of artificial intelligence"	Q1 2025 depreciation up $217M, mostly AWS	10-Q Q1 2025
Oracle	six years	FY2026 depreciation $7.6B, from $3.9B	10-K FY2026
CoreWeave	six years	raised from five on 1 Jan 2023; $20M expense reduction that year	10-K 2025

The honest reading is that the accounting is defensible and the economics are uncertain, which are different statements. Amazon is the only one that moved the other way, and it cited AI. Meta's extension alone added $2.59 billion to 2025 net income. CoreWeave went to six years in January 2023, before it had an IPO or a rated loan, which is at least consistent. The evidence for long lives is real: the A100 contract into 2029, Hopper and Ampere fleets "largely sold out", an index that fell to $1.70 and recovered to $2.59, and a three-year-old H100 system reselling at about 45 percent of new, per Yahoo Finance in December 2025. The evidence against is also real: the B200 rents for 2x an H100, and an H100 bought at 2023 prices to earn $8 an hour now earns $2.59, which suits a six-year straight line only if the purchase was underwritten at $2.59. The useful life that matters is the ratio of the rent an asset earns in years four to six to the debt service sized against years one to three, and no filing lets an outsider compute that per cluster.

Own or rent: the arithmetic

The depreciation debate is the own-versus-rent decision written in GAAP, which is why it matters to someone my size. Here is the calculator, every assumption a named input, because the answer flips on utilisation and assumed life and little else. The defaults are placeholders, not quotes.

#!/usr/bin/env python3
# breakeven.py: own-vs-rent calculator for a GPU server. Python 3.11+, stdlib only.
# All inputs are YOUR assumptions. The defaults below are placeholders to show the
# shape of the answer, not market quotes. Change them before believing the output.
from dataclasses import dataclass

@dataclass
class Inputs:
    server_price_usd: float = 250_000   # delivered 8-GPU server incl. networking share
    gpus_per_server: int = 8
    useful_life_years: float = 6.0      # CoreWeave/Microsoft/Google/Oracle book 6; Amazon 5; Meta 5.5
    residual_fraction: float = 0.10     # resale value at end of life as fraction of price
    interest_rate: float = 0.08         # blended cost of capital (DDTL 4.0 was ~5.9% fixed; DDTL 5.5 SOFR+5.5)
    server_power_kw: float = 10.2       # IT load per server at full tilt
    pue: float = 1.3                    # facility overhead multiplier
    electricity_usd_per_kwh: float = 0.08
    colo_usd_per_kw_month: float = 150  # space, cooling, remote hands, per kW of IT load
    opex_fraction_of_capex_per_year: float = 0.03  # support contracts, spares, staff share
    utilisation: float = 0.70           # fraction of 8,760 h/yr actually billed or used
    rental_usd_per_gpu_hour: float = 2.59  # what you would pay instead (OCPI-H100, 4 Oct 2026)

def own_cost_per_gpu_hour(i: Inputs) -> dict:
    hours_per_year = 8_760
    depreciable = i.server_price_usd * (1 - i.residual_fraction)
    depreciation_yr = depreciable / i.useful_life_years
    # simple interest on average outstanding balance over the life
    interest_yr = i.server_price_usd * i.interest_rate / 2
    power_kwh_yr = i.server_power_kw * i.pue * hours_per_year * i.utilisation
    power_yr = power_kwh_yr * i.electricity_usd_per_kwh
    colo_yr = i.server_power_kw * i.colo_usd_per_kw_month * 12
    opex_yr = i.server_price_usd * i.opex_fraction_of_capex_per_year
    total_yr = depreciation_yr + interest_yr + power_yr + colo_yr + opex_yr
    billed_gpu_hours = i.gpus_per_server * hours_per_year * i.utilisation
    return {
        "depreciation_per_gpu_hr": depreciation_yr / billed_gpu_hours,
        "interest_per_gpu_hr": interest_yr / billed_gpu_hours,
        "power_per_gpu_hr": power_yr / billed_gpu_hours,
        "colo_per_gpu_hr": colo_yr / billed_gpu_hours,
        "opex_per_gpu_hr": opex_yr / billed_gpu_hours,
        "total_per_gpu_hr": total_yr / billed_gpu_hours,
    }

def breakeven_utilisation(i: Inputs) -> float:
    # utilisation at which owning costs the same per billed GPU-hour as renting
    lo, hi = 0.01, 1.0
    for _ in range(60):
        mid = (lo + hi) / 2
        j = Inputs(**{**i.__dict__, "utilisation": mid})
        if own_cost_per_gpu_hour(j)["total_per_gpu_hr"] > i.rental_usd_per_gpu_hour:
            lo = mid
        else:
            hi = mid
    return hi

if __name__ == "__main__":
    base = Inputs()
    for life in (3.0, 6.0):
        for util in (0.5, 0.7, 0.9):
            i = Inputs(**{**base.__dict__, "useful_life_years": life, "utilisation": util})
            c = own_cost_per_gpu_hour(i)
            print(f"life {life:.0f}y util {util:.0%}: own ${c['total_per_gpu_hr']:.2f}/GPU-h "
                  f"(dep {c['depreciation_per_gpu_hr']:.2f}, int {c['interest_per_gpu_hr']:.2f}, "
                  f"power {c['power_per_gpu_hr']:.2f}, colo {c['colo_per_gpu_hr']:.2f}) "
                  f"vs rent ${i.rental_usd_per_gpu_hour:.2f}")
    for life in (3.0, 6.0):
        i = Inputs(**{**base.__dict__, "useful_life_years": life})
        be = breakeven_utilisation(i)
        print(f"life {life:.0f}y: break-even utilisation vs ${i.rental_usd_per_gpu_hour}/h rent = "
              f"{'>100%' if be >= 0.999 else f'{be:.0%}'}")

I ran it as written in a scratch directory on 5 October. With the placeholder inputs, a server depreciated over six years costs $1.63 per GPU-hour to own at 70 percent utilisation and $2.23 at 50 percent, against $2.59 transacted rent; over three years, $2.39 and $3.30. Break-even utilisation against $2.59 is 43 percent on a six-year life and 64 percent on three. Those are arithmetic on my assumptions, not market facts; the point is the shape. Depreciation is the largest line at every utilisation, interest is a fifth of it, and power at eight cents is nearly a rounding error. That is SemiAnalysis's 2023 finding, that GPU clouds are a capital business, and it is why the industry argument has collapsed into one accounting estimate. It is also why six-year books are not obviously wrong for hyperscalers running at internal utilisation. The three-year case is the neocloud that bought at 2024 prices, filled half its hours, and now competes with Nebius's $0.79 preemptible tier.

How someone my size actually buys

I do not own GPU servers, and the calculator is why: my accelerator load is bursty, and nothing beats a marketplace or a preemptible tier for bursty load. What changed in 2026 is that the options became legible. A transacted index exists, so I know when a quote is 50 percent over market. Marketplaces publish machine-readable offers: the Vast.ai query in the table is one unauthenticated GET a scheduler can call before placing a job. Nebius and CoreWeave publish spot prices at a third to a fifth of on-demand, and Together sells 180-day reservations 20 percent under list without a sales call. For sustained load, the sellers' filings say how to negotiate: CoreWeave says shorter contracts carry "a higher ASP and margins", and Nebius says 70 percent of its deals include prepayments covering 50 to 60 percent of the capex. A buyer who commits a year and prepays is financing the seller's server and should be priced like a lender, not a tenant.

The discipline this demands is the one from what scale to zero cost me: utilisation is the only variable in the calculator a small operator controls. Checkpointing a job so it can move to a cheaper preemptible instance, bin-packing inference onto fewer cards, and sleeping idle capacity are not optimisations at the edge of the cost; at 50 percent utilisation they are the cost. The inference serving stack and the GPU bill are one problem seen from two ends.

What I take from the quarter is that the GPU cloud has stopped being a shortage story and become a credit story, and credit stories are decided by cash flows in years four to six, not press releases in year one. Rent recovered from $1.70 to $2.59 and list prices rose, good for every six-year balance sheet and bad for anyone who assumed hours would keep getting cheaper. Debt got cheaper and longer, from 15 percent DDTLs to a 5.9 percent A3 facility, which is also good until an offtaker that is 36 percent of revenue renegotiates. The useful-life argument will be settled by the H100 rental price in 2028, not by anyone's model; I would bet on the A100-to-2029 contract being more typical than a two-year life, noting I am a buyer who benefits if I am wrong. This quarter I will keep renting, wire the Vast.ai and Nebius spot prices into the job scheduler so placement follows the index, and move my one sustained GPU workload onto a prepaid 180-day reservation, because the filings say that is the cheapest money in the market and the sellers have explained exactly why.

---

Related: What Scale to Zero Cost Me to Build, AI Power in 2026 and FinOps for AI Token Costs.