A $45B Compute Deal, the Neocloud Chill, and a Prospectus Full of $750B: The Truth Behind Anthropic's Compute Arms Race
A $45B Compute Deal, the Neocloud Chill, and a Prospectus Full of $750B: The Truth Behind Anthropic’s Compute Arms Race
Introduction
It is October 21, 2026. Open any AI-industry news feed and you’ll see a strange picture of “scorching and freezing.” On the one hand, Anthropic is negotiating a $45-billion compute deal with Nscale, with a signature possible within days. On the other, a cluster of new-style cloud providers that call themselves “Neocloud” are taking a beating all at once — CoreWeave lost more than 23% of market value in eight days, Nebius fell 32% in two weeks, Firmus pulled its $5-billion IPO. Just as the sector grows jittery, Nscale is leaning on that Anthropic order to hold its footing at the door of a ~$30-billion-valuation IPO.
Even more telling, Anthropic’s still-unpublished prospectus draft — already widely reported — is laying its next-7-to-10-year compute bill out for everyone: total committed contracts around $450B now, potentially another $150B on top of what’s being negotiated, heading toward $600B or even $750B. Booking a net loss of $80.7B in half a year (of which ~$72B was an accounting mark of equity instruments) while building this kind of compute empire — is that honest investment or drinking poison to quench thirst?
In this article I want to unpack the truth behind Anthropic’s compute arms race: starting with this $45B order itself, then examining why the Neocloud business model swings so hard between hot and cold, then reconstructing an AI giant’s real infrastructure ledger from the numbers in the prospectus, and finally leaving a few questions worth a calm second thought — for both engineers and investors.
Part One What a $45 Billion Order Actually Buys
Let’s first get the order straight.
Per Bloomberg, Zhitong Finance and other outlets citing sources, Nscale is in talks to sell Anthropic roughly $45 billion worth of compute, at the second building of Nscale’s Monarch data-center campus in Mason County, West Virginia. Key parameters: this second-phase building can provide about 460 megawatts of computing capacity, powered by Nvidia’s newest-gen Vera Rubin AI chips; Anthropic also holds a preferential right to lease the campus’s planned third building. Per its IPO filing, Anthropic had already signed an infrastructure agreement worth up to about $44.6 billion; this $45B is an “additional bet” centered on the second building, possibly closing within days. Talks continue; details may change.
A few points are easy to conflate here, so let’s be careful.
First, this $45B is not a one-time payment; it is a multi-year lease commitment. Per the earlier Bloomberg report of late August, Anthropic agreed to pay Nscale $45 billion over six years for 460 MW of capacity at Monarch — about $7.5B a year, or ~$625M a month. So it buys not “compute on the shelf today” but “new capacity that comes online only in late 2027 / early 2028.”
Second, the Vera Rubin chips at issue haven’t shipped at commercial scale yet. This is Nvidia’s seventh-generation platform. A single NVL72 rack packages 72 Rubin GPUs plus 36 Vera CPUs into one fully liquid-cooled assembly, delivering 3.6 ExaFLOPS of NVFP4 inference throughput and 2.5 ExaFLOPS of training throughput; a full 40-rack Vera Rubin POD scales to 1,152 Rubin GPUs and an aggregate 60 ExaFLOPS. In other words, Anthropic signed a contract betting that future supply meets future demand — the same calculus behind nearly every entry in its compute ledger.
Third, this campus is unusual. Monarch is billed as the United States’ first state-certified AI microgrid campus, spanning 2,250 acres with total power scalable beyond eight gigawatts. Its power model is extreme: 864 Caterpillar G3500 reciprocating natural-gas generators on site (plus ~120 backup units), totaling ~2.16 GW of generation, designed to operate independently of the local grid. That also means the campus’s carbon footprint is set by gas consumption, not by renewable-energy certificates; a West Virginia law (HB 2014, signed 2025) cleared the way for the microgrid and state-level planning. Environmental groups including the Sierra Club’s WV chapter have flagged concerns.
Put those three together: what Anthropic’s $45B really buys is an insurance policy that locks in future-years capacity, on the newest chips, with on-site energy.
Part Two The Neocloud Chill: Hot Orders, Cold Stock Prices, Heavy Debt — and Why It Matters
But behind this big order, the Neocloud track is experiencing a violent reversal. To understand why this order is so important, you need to see that reversal.
The numbers first. On Oct 9, Firmus withdrew a $5B IPO. On Oct 18, the UK’s Telegraph reported Microsoft would cancel a collaboration with Nscale (that had been part of Monarch’s original plan — Microsoft signed a letter of intent in March but exited the West Virginia site over the summer, and Anthropic stepped in). That nearly pushed Nscale’s IPO off a cliff. Then from Oct 12–19, multiple listed Neoclouds were sold off: CoreWeave -23.91%, Nebius -18.96%, Core Scientific -8.59%. CoreWeave’s own public figures are starkest: total liabilities grew from $5.109B in 2023, to $45.967B by end-2025, to $72.046B (≈RMB 481.2B) by end-June 2026.
Gartner added another jab: an Oct 15 report asserted that by 2030, 80% of Neoclouds will stop operating independently — merged away or gone. Analysts also estimate only about 35% of North America’s announced data-center capacity will actually be built.
Why is this industry so fragile? Deep down, the Neocloud business model has an unavoidable core: first bind a big-model client with a long-term compute framework agreement, pledge the order and GPUs as collateral to borrow from private-credit lenders, then lock Nvidia GPU allocation and build an AI computing center. This “pledge-orders, borrow-to-expand” model shines in an up-cycle but piles three mountains onto itself by nature:
- Cash flow and financing costs: multi-year lease-style commitments mean revenue can only be recognized after capacity comes online, yet debt interest is due today;
- GPU depreciation: chips iterate fast (Grace Blackwell → Vera Rubin), so assets can lose value before they’ve run their course;
- Supply-demand mismatch: once compute supply exceeds real demand, price wars and vacancy break out simultaneously.
That’s exactly why, when a big client like Anthropic signs a $45B order and becomes Nscale’s largest customer, it’s not just revenue — it’s a ballast stone to “endorse” an IPO valuation. Even though most of that ballast still sits in the “contracted but not yet earned” column.
A telling comparison: Nscale’s filing shows recognized revenue of only ~$33M for all of 2025, ~$37M in Q1 2026, and just over $100M in Q2 2026. In other words, the $45B contract value is just a number on a future order book; real revenue only starts climbing when Monarch Phase 1 ramps in late 2027. Treating the order book as revenue, and valuation expectations as reality — that’s one of the psychological roots of this Neocloud panic.
Part Three The Ledger in the Prospectus: $600B or $750B?
If the $45B order is the tip of the iceberg, the compute ledger in Anthropic’s prospectus draft is the whole iceberg. Let’s lay it out.
Per multiple reports, for the six months ending June 30, Anthropic’s revenue was ~$16.3B (versus just $1.1B a year earlier), but its net loss was $80.7B — of which about $72B was an accounting fair-value change on convertible notes and similar instruments, a book swing rather than an operating loss. In the same period it spent nearly $9B in cash on compute and infrastructure (versus $2.3B in the year-ago period), and in just the first half it spent more on compute than it did in all of 2025 ($7.3B).
The future commitments are the scarier part. The draft shows that over the next 7–10 years, existing compute purchase agreements total roughly $450B in contracted value, and most cannot be cancelled. Breaking it down:
Anthropic Compute Vendor Ledger (multiple sources, USD$B)
┌──────────────────┬──────────────────────────────────────┐
│ Vendor │ Commitment / Note │
├──────────────────┼──────────────────────────────────────┤
│ Google Cloud │ ~111 │
│ AWS │ ~110 │
│ Microsoft │ ~31 │
│ Nscale + Lambda │ up to ~86 total (Nvidia chips, 33-34)│
│ Google chips(Bogru)│ >160 lease obligation via Broadcom │
│ Other clouds │ several │
│ Warburg Monarch │ ~45 (in talks, 2nd building) │
├──────────────────┼──────────────────────────────────────┤
│ Existing total │ ≈450 (contracted, mostly non-cancel) │
│ Negotiating add │ up to another ~150 │
│ Total infra cost │ already >600; ceiling ~750 │
└──────────────────┴──────────────────────────────────────┘
A few points here deserve a pause.
First, it spreads eggs across a basket of “multi-vendor + multi-chip-generation + cross-border + multi-financing-structure.” It binds AWS, Google Cloud, and Microsoft — the three hyperscalers — and deeply bets on more peripheral compute suppliers like Nscale, Lambda, and SpaceX’s Colossus, while taking on more than $160B in Google TPU lease obligations via a Broadcom deal. Walking on two legs and sprinting on many is the portrait of this infrastructure system.
Second, most are rigid “non-cancellable” commitments. That means even if future model demand disappoints, this money largely must keep flowing out. That’s why the market keeps wondering whether the compute arms race is overdone.
Third, much of the capacity is “forward futures.” Vera Rubin hasn’t shipped at scale; Monarch won’t start delivering until late 2027, yet contracts are signed. You’re buying the future, betting the demand persists.
Here’s a small script to visualize the cash-flow mismatching of this kind of “forward capacity” — heavy spending early, revenue ramping only after commissioning:
def capacity_cashflow(contract_amts, pay_years, online_year_start):
years = list(range(2026, 2036))
rows = []
for name, total in contract_amts.items():
annual_out = total / pay_years # committed/yr (in $B)
for y in years:
if y < online_year_start[name]:
recognized, spent = 0.0, annual_out
else:
ramp = min(1.0, (y - online_year_start[name] + 1) / 3.0) # 3yr ramp
recognized, spent = total * ramp, annual_out
rows.append((name, y, round(spent,1), round(recognized,1)))
return rows
contracts = {
"Google": 111.0, "AWS": 110.0, "Microsoft": 31.0,
"Nscale+Lambda": 86.0, "Monarch(talks)": 45.0,
}
online = {"Google":2027, "AWS":2027, "Microsoft":2027,
"Nscale+Lambda":2027, "Monarch(talks)":2028}
cashflow = capacity_cashflow(contracts, pay_years=10, online_year_start=online)
from collections import defaultdict
by_year = defaultdict(lambda: [0.0, 0.0])
for _name, y, spent, rec in cashflow:
by_year[y][0] += spent; by_year[y][1] += rec
print("Year Annual spend($B) Recognized value($B)")
for y in sorted(by_year):
s, r = by_year[y]
print(f"{y} {s:>16.1f} {r:>20.1f}")
(Note: this is a teaching-level simplification, meant to show the “heavy early spending, revenue ramp later” structure intuitively. Real figures are subject to Anthropic’s prospectus and contract terms.)
What this code conveys is not a prediction of Anthropic’s cash flow, but a way to understand why this class of forward-capacity contract keeps being re-priced by the market — because the certainty of spending is far higher than the certainty of revenue realization.
Part Four From “One-Machine Decisiveness” to Anthropic’s Compute Philosophy
Before digging deeper into the numbers, there’s a tactical observation worth expanding: why is Anthropic locking in compute so aggressively in advance? Behind it is a clear-cut “compute philosophy.”
It treats compute as a strategic variable almost as important as model capability. For a company whose training and inference both depend heavily on GPUs, compute is “ammunition.” Anthropic chose to sign contracts while Vera Rubin hasn’t shipped at scale and Monarch won’t come online until late 2027 — essentially not because it “needs compute now,” but because it “predicts a future shortage.” Locking in early means occupying supply-chain seats before rivals scramble for capacity. It mirrors hyperscalers hoarding storage around 2014 and miners grabbing GPUs in the 2021 mining wave — a classic “pre-position on a scarce resource” strategy.
Its positioning is “multi-vendor hedging” rather than “one big bet.” It hugs the three hyperscalers (AWS, Google Cloud, Microsoft), bets on upstarts like Nscale, Lambda, and SpaceX, and takes on Google TPU lease obligations via Broadcom. On the surface this looks like “casting a wide net,” but it implies a risk-hedging logic: different vendors’ delivery timing, chip generations, and power conditions differ, so running multiple tracks both diversifies the risk of any single supply chain slipping and enables rapid switching if one chip generation underperforms. In other words, Anthropic is buying not just compute, but the “ability to hedge against supply-chain uncertainty.”
It treats “forward futures” as a tool for growth certainty. For capital markets, a 7-to-10-year contract commitment is the strongest endorsement of “future demand certainty”; for a model company, it’s the confidence that “capacity keeps up no matter how demand fluctuates.” Together, they let Anthropic use a string of non-cancellable deals, on the eve of its IPO, to send the market a strong signal: “I’m bullish on the future, and I’ve locked supply.”
Of course, the other side of this philosophy is obvious: certainty is a double-edged sword. The more rigid the commitment, the less room to maneuver. Once there’s a technological pivot (say, no longer relying on large-scale pretraining) or downstream demand misses, these non-cancellable bills turn from “competitive moat” into “financial baggage.” That’s exactly why the market, while giving high valuations, keeps flagging downside risk.
Here’s a code sketch simulating how “multi-vendor hedging” smooths single-point risk when some vendors deliver late:
package main
import "fmt"
// vendor represents one compute supplier's delivery plan
type vendor struct {
name string
contracted float64 // contracted capacity (MW)
deliverRate float64 // actual delivery rate (0~1)
}
// totalDelivered computes expected delivery of a multi-vendor portfolio
func totalDelivered(vs []vendor) float64 {
sum := 0.0
for _, v := range vs {
sum += v.contracted * v.deliverRate
}
return sum
}
func main() {
vendors := []vendor{
{"AWS", 1200, 0.9},
{"GoogleCloud", 1000, 0.88},
{"Nscale(Monarch)", 460, 0.35}, // new microgrid campus, high risk
{"Lambda", 260, 0.7},
{"SpaceXColossus", 450, 0.8},
{"Bogru(TPU)", 300, 0.75},
}
need := 2600.0
got := totalDelivered(vendors)
fmt.Printf("need=%vMW portfolio=%.1f coverage=%.1f%%\n",
need, got, got/need*100)
only := []vendor{{"Nscale(Monarch)", 460, 0.35}}
fmt.Printf("if only betting Monarch: expected=%vMW (far below need)\n",
totalDelivered(only))
}
(Example numbers are to demonstrate the multi-vendor hedging logic; agreements prevail.)
What this code conveys: Anthropic’s multi-track layout keeps the portfolio’s “expected delivery” from collapsing even if one high-risk new campus delivers only 35% — the mature suppliers still backstop it. That’s one technical reason it dares to back thousands of billions in commitments.
Part Five Three Cool-Headed Questions Behind the Blazing-and-Freezing Picture
Putting the $45B order, the Neocloud chill, and the prospectus ledger together, I want to leave three questions worth independent thought.
First, honest investment or overbuilt arms race? Anthropic’s willingness to back thousands of billions in commitments is, logically, a bet that future AI demand will long persist, and it uses early capacity lock-in to buy growth certainty. Conversely, Gartner’s “80% of Neoclouds won’t survive 2030” and Bernstein’s “35% build-to-complete” remind us: if supply really exceeds demand, this “fuel” may become “baggage.” Giants can absorb a lot, but no one is immune to the cycle.
Second, can cost “come ashore”? SenseTime just open-sourced RoboRSI, pushing embodied-task cost to 1/50 of GPT-6-Astra; DeepSeek and others keep pressing unit-task cost down with extreme engineering. As “doing the same job well” gets progressively cheaper, the “enormous compute” that giants pay billions for may no longer equal “absolute dominance.” The dividend of the arms race is flowing more and more toward “efficiency” than “sheer scale.”
Third, the tug between prospectus and valuation. On one hand Anthropic posted Q3 revenue of $16.8B, adjusted profitability, and positive free cash flow — a signal that “self-sustaining capacity” has turned positive. On the other, slowing monthly revenue growth and investor concerns about cash burn have pulled IPO valuation expectations from ~$2T down to the ~$1.5T range. The market isn’t denying Anthropic’s strength; it’s cooling the heat of the “burn to grow” narrative.
Conclusion
From a West Virginia natural-gas microgrid, to Vera Rubin’s NVL72 racks, to a prospectus full of breath-taking $600B/$750B numbers — Anthropic is strapping itself to a fast-spinning arms machine with a string of astronomical compute commitments. This $45B order is both the anchor propping up Nscale’s IPO valuation and the insurance it buys for its own growth certainty.
But for everyone whose pulse jumps at those numbers, the real homework may be: in the world of the compute arms race, learn to tell “order book” from “revenue,” “commitment” from “delivery,” “scale” from “efficiency.” The giants are betting an unprecedented amount of capital on a future — and whether that future turns out sweet or bitter depends, in the end, not on who bets the most, but on whose ledger actually delivers when it’s supposed to.