Everything About AI
NextFin News - A two-year-old London company with a $1 billion half-year loss has filed to go public in New York at a valuation that would make it one of the largest technology listings of the decade. Nscale's S-1, filed on September 18 with a $103 billion contract backlog, is doing more than testing IPO demand: it is forcing Wall Street to decide whether AI data-center specialists are durable infrastructure owners or highly leveraged middlemen standing between Nvidia's chips and the hyperscalers that buy them.
The filing lands at a moment when the neocloud category has moved from niche to central. Synergy Research Group estimated that neocloud revenue reached $9 billion in the fourth quarter of 2025 alone, up 223% year over year, and exceeded $25 billion for the full year, with the market on track to approach $400 billion by 2031. Nscale is now asking public investors to underwrite the next leg of that growth — with financials that show just how much cash must be burned before a backlog becomes an income stream.
The Filing: A $103 Billion Backlog Built on $140 Million of Revenue
Nscale Limited filed a registration statement with the Securities and Exchange Commission for a proposed listing on the New York Stock Exchange under the ticker NSCL. Goldman Sachs, J.P. Morgan and Morgan Stanley are the lead bookrunners, supported by a 23-bank syndicate. The number of shares and the price range have not yet been set, but reporting on the company's pre-IPO financing points to a valuation conversation in the $30 billion to $35 billion range — more than double the $14.6 billion mark set in a March 2026 Series C round.
The numbers inside the prospectus are what stopped traders. Revenue for the first six months of 2026 was $140.6 million, up 1,252% from $10.4 million a year earlier. The net loss for the same period was $1.02 billion, up from a $368.9 million loss in the first half of 2025. Cost of revenue was $189.6 million, meaning the company did not make a gross profit: it cost more to deliver the service than customers paid. In 2024 the company did show a gross profit of $6.3 million on revenue of $19.1 million; since then, as the buildout accelerated, losses have widened.
Management wrote in the prospectus that it once had "substantial doubt" about whether the company could continue as a going concern — the accounting phrase that sits one step short of bankruptcy risk — before concluding that it could defer, reduce or cancel capital spending if financing failed to arrive. The doubt has since been backstopped: on September 15 the company received $3.1 billion from Nvidia in exchange for unsecured convertible loan notes. Days earlier, reports said Nscale was negotiating a $3.5 billion pre-IPO round split between roughly $1.5 billion of convertible notes led by hedge fund Third Point and about $2 billion coming directly from Nvidia.
Management identified that the Company's forecast funding requirements included reliance on uncommitted debt and equity financing, which initially raised substantial doubt about the Company's ability to continue as a going concern.
Then there is the backlog, the single asset asked to justify the valuation. Nscale says total contracted revenue reached $103 billion after a six-year agreement with Anthropic signed on August 25, valued at up to approximately $44.6 billion. Microsoft and Anthropic together account for $87.7 billion — roughly 85% of the total. But of that $103 billion, only $2.6 billion was "active" as of the end of August. The remainder is scheduled across years: of the $34.5 billion in remaining performance obligations recorded at the end of 2025, the company expects to recognize about 16% in the next two years, 40% in the following two, and the rest over the three years after that.
In other words, the market is being asked to value promises that arrive on a seven-year clock, while the costs of keeping those promises land on this year's income statement. That mismatch is the entire investment case — and the entire risk.
What the IPO Is Really Pricing: Execution, Not Assets
The first question the market must answer is whether Nscale is selling compute or selling contracts. A backlog is not revenue. It becomes revenue only if the company can build the data centers, secure the power, install the GPUs and keep the systems running well enough to invoice under take-or-pay terms. The neocloud model is capital expenditure first, revenue recognition years later. Nscale's negative gross margin in the first half of 2026 is the mechanical signature of that sequence: the costs hit the income statement now, while the contracted cash flows sit in the future.
This is why the $30 billion to $35 billion valuation range is a bet on execution rather than on assets already producing. The closest public proxy is CoreWeave, which priced its own IPO at $40 a share on March 27, 2025, raised $1.5 billion, and debuted at $39 for a valuation of roughly $23 billion — after first targeting a $47 to $55 range and then downsizing the offering on weak demand. A year and a half of trading later, CoreWeave's shares have ranged from a low of $60.55 to a high of $183.58 and were changing hands near the mid-$80s in late September, a market capitalization of roughly $40 billion. CoreWeave reported second-quarter 2026 revenue of $2.575 billion, up 112% year over year, against a revenue backlog of about $104 billion and adjusted EBITDA margins above 50%.
Nscale's backlog is essentially the same size as CoreWeave's, but its trailing revenue run-rate is a small fraction of CoreWeave's, and it is not yet EBITDA-positive. The market is being asked to value Nscale's backlog at parity with CoreWeave's while applying a discount for the fact that almost none of it is earning money. That is a demanding comparison, and it is the reason the final price range in the next S-1 amendment will matter more than the headline backlog number.
The Nvidia financing deepens the ambiguity. A $3.1 billion convertible note from the company whose GPUs Nscale rents out is both validation and dependency. It is validation because Nvidia would not commit at that scale to a counterparty it did not expect to survive and build. It is dependency because Nscale's entire value proposition rests on access to Nvidia's supply — and its largest customer, Microsoft, is simultaneously pushing Nvidia's competitors and designing its own AI accelerators.
Neoclouds Are Not Just Middlemen — But the Margin Has to Prove It
The neocloud thesis is that AI workloads are different enough from general cloud computing to justify a specialized layer. Training runs and large-scale inference need dense GPU clusters, high-bandwidth networking and power availability that general-purpose clouds were not built to prioritize. The growth data supports the thesis: Mordor Intelligence estimates the neocloud market at $35.22 billion in 2026, growing to $236.53 billion by 2031 at a 46% compound annual rate, with AI and machine-learning workloads projected to grow fastest of all.
The question is who captures the economics. A pure middleman that rents GPUs and sublets them earns the spread between its financing and power cost and the price its customers will pay. That spread compresses when GPU supply loosens, when hyperscalers build their own capacity, or when customers negotiate harder on the back of concentrated buying power. Nscale's concentration risk is not theoretical: Microsoft and Anthropic represent $87.7 billion of $103 billion in contracts. If either renegotiates, delays, or fails to activate capacity, the backlog — the single asset justifying the valuation — shrinks.
There is, however, a structural reason neoclouds may hold more pricing power than the middleman label suggests. Power and grid connections are the binding constraints, not chip allocation alone. Nscale's flagship campus in Narvik, Norway, was developed with industrial investor Aker, and the company has signed long-term renewable power purchase agreements. A data center that already has a grid connection and a power contract is a scarce asset in a market where hyperscalers wait years for capacity. That scarcity is what can convert a GPU reseller into an infrastructure owner. The S-1 reflects the ambition: Nscale describes itself as a "full-stack AI cloud platform" that brings together software, compute and power, not merely a hosting operator.
Founded in May 2024 by Josh Payne and Nathan Townsend and spun out of the crypto-mining and renewable-infrastructure business Arkon Energy, Nscale has raced from Bitcoin mining economics to AI compute in roughly the time it takes a traditional utility to permit a single substation. That speed is the appeal. It is also the execution risk that the public market will now price.
The Second-Order Trade: Nvidia's Balance Sheet Is the Real Benchmark
The second-order implication of the Nscale filing is that it is less a test of Nscale than a test of Nvidia's credit judgment. When a chip supplier takes a $3.1 billion convertible position in a customer, the market begins pricing the supplier's exposure to that customer's survival. Nvidia has done this before — it contributed a $250 million order to CoreWeave's IPO — but the scale here is an order of magnitude larger, and it arrives while Nvidia's own revenue is already priced on the assumption that AI infrastructure buildout continues uninterrupted.
The transmission chain runs: Nscale IPO prices → neocloud sector re-rated → Nvidia's customer-concentration risk repriced → discount rates on AI capex adjust. If Nscale prices at the top of its range and trades up, it confirms that investors are willing to pay for backlog quality, and the entire neocloud complex — CoreWeave, Lambda, Crusoe, and the power and data-center companies that supply them — earns a higher multiple. If the deal is downsized or prices weakly, the first casualty is not Nscale; it is the assumption that every announced AI compute contract is bankable at face value.
That is the already-priced conventional wisdom this IPO challenges. The market has treated AI infrastructure announcements as additive and durable. Nscale's filing forces a distinction between signed contracts and activated capacity, and between gross backlog and collectible revenue. The gap between the two is where the risk premium lives.
The Counter-Thesis: A Cyclical Arbitrage That Will Self-Correct
The strongest argument against the structural view is that neoclouds are a cyclical arbitrage that will correct itself. Hyperscalers are not passive customers; they are the largest capital spenders in the industry, and they have every incentive to internalize capacity once they understand the workload. Microsoft designs its own AI accelerators and maintains multiple cloud suppliers. When GPU supply catches up with demand — as it tends to after every semiconductor upcycle — the scarcity rent that neoclouds extract from chip allocation disappears. CoreWeave's stock, down more than 50% from its June 2025 peak, is the market's own evidence that the trade can mean-revert quickly.
This counter-thesis has force. It is backed by the observable pattern of every infrastructure buildout: margins are highest when capacity is scarce and compress as supply arrives. The specific falsifying signal for the structural case is this: if CoreWeave's adjusted EBITDA margin compresses by more than 15 percentage points over two consecutive quarters while GPU supply normalizes, the neocloud model is a cyclical spread business rather than a structural moat, and Nscale's backlog should be discounted accordingly.
What Comes Next
The near-term read is straightforward: the Nscale IPO is a liquidity event for early backers and a sector-wide referendum. A strong pricing would reopen the IPO window for other AI infrastructure names and lift the entire neocloud complex. A weak one would force a repricing of backlog multiples across CoreWeave, Lambda and Crusoe, and could tighten the convertible financing that these capital-intensive businesses rely on.
The medium-term question is execution. Nscale must convert $2.6 billion of active backlog into revenue while building out the remaining commitments. The 16/40/remainder recognition schedule in the S-1 implies most of the value arrives in years three through seven. That is a long runway for execution risk, customer renegotiation and financing-cost changes.
The long-term structural call is that AI compute demand is durable, but the winners will be determined by power and capital access, not by GPU allocation alone. Nscale's Norway campus, its power purchase agreements and its vertical integration into software and orchestration are attempts to own the scarce inputs rather than rent the abundant ones. If that works, neoclouds are not middlemen — they are the new utilities. If it does not, they are well-financed distributors whose margins revert to their cost of capital.
What to watch: the final IPO price range in the next S-1 amendment; the first quarterly revenue recognition against the $2.6 billion active backlog; CoreWeave's adjusted EBITDA margin trend as the public proxy; and whether Microsoft and Anthropic activate capacity on schedule. The single signal that would break the structural thesis is a backlog contraction or a material renegotiation by either anchor customer within the next two quarters.
The market is not paying for Nscale's data centers. It is paying for the belief that a $103 billion promise can be turned into a $103 billion income stream — and that is a much harder engineering problem than building the racks.
Explore more exclusive insights at nextfin.ai.
