Nscale, a UK cloud computing newcomer backed by NVIDIA (NVDA) and holding contract commitments worth hundreds of billions of dollars, is facing the same predicament as its Australian peer Firmus did just a month ago: investors are demanding an IPO valuation cut nearly in half, and Microsoft has announced its withdrawal from a key data center project.
According to Bloomberg, Nscale had previously sought to list at a valuation of approximately $30 billion, but a group of investors has made clear that $16 billion to $19 billion is "more likely to win buyer approval" — a reduction of more than 40% from the originally targeted valuation. Meanwhile, Reuters, citing The Telegraph, reported that Microsoft has terminated its cooperation agreement with Nscale for a data center campus in Loughton, Essex, UK, because National Grid officials informed Nscale that the site may not receive power supply until as late as the 2030s. Microsoft declined to comment.
This marks the fifth case in less than a month of an AI data center listing plan hitting setbacks, following four previous cases involving SoftBank's SB Energy, Firmus, Philippine PLDT subsidiary Vitro, and Blackstone's AirTrunk. Goldman Sachs' credit team attributed this series of events to "AI trades encountering the cost of capital," while the credit market had already sounded the alarm weeks earlier through deteriorating bond prices and credit default swaps.
Microsoft Withdraws: No Power in Sight, Anchor Customer Leaves First
The failure of the Loughton project reflects the core contradiction in current AI infrastructure construction: contracts can be signed in a single day, but grid connections may be indefinitely delayed.
According to Reuters, the direct trigger for Microsoft's termination was National Grid's notification that the site has no prospect of receiving power supply until the 2030s. Nscale responded that its "relationship with Microsoft remains solid and long-term," characterized the issue as a "grid problem within London," and said it is "exploring alternative energy solutions." However, with the anchor customer already gone, such statements can hardly compensate for the substantial commercial loss.
This is not the first time this year that Microsoft has chosen to withdraw from a third-party AI computing project — according to previous Bloomberg reporting, in June Microsoft declined a $3 billion Oracle cloud computing contract. When the hyperscaler with the most abundant cash reserves globally begins to be selective about renting computing power, the core business logic of new cloud companies that pitch "bringing computing power online faster than anyone" is directly challenged.
The situation at the Norwegian flagship campus is equally discouraging. Nscale's Glomfjord facility north of the Arctic Circle has been delayed by six months due to cooling system leaks, local regulations, construction delays, and fire incidents. Of the planned 30 megawatts of capacity, only 6.4 megawatts is currently in actual operation.
The Real Ledger Behind Billion-Dollar Contracts: A 4% Conversion Rate and a 107x Valuation Bubble
Nscale's prospectus looks impressive on headline numbers, but upon closer examination, the data reveals a massive gap between reality and promises.
The company disclosed a total contract value of up to $103.4 billion, a dramatic expansion from approximately $100 million when it was founded two and a half years ago. However, the actual value of contracts in operation is only $2.6 billion; deployed capacity stands at 55 megawatts (of which 48 megawatts is leased colocation), while signed IT capacity commitments total approximately 1,346 megawatts — an actual conversion rate of only about 4%. Microsoft (contract ceiling of $43.8 billion) and Anthropic (contract ceiling of $44.6 billion, both from a West Virginia campus that has yet to secure financing) together account for approximately 85% of the order backlog, while the Anthropic contract terms allow it to exit without additional obligations, and the related financing has not yet been secured.
On the financial side, Nscale reported revenue of $140.6 million for the first half of 2026, with a net loss of $1.02 billion — meaning it loses approximately $7 for every $1 of revenue. Annualizing first-half revenue, roughly $280 million in annual revenue corresponds to a $30 billion valuation, implying a price-to-sales ratio of approximately 107x. Even at the investor-accepted range of $16 billion to $19 billion, the price-to-sales ratio would still be between 57x and 68x.
Additionally, the Financial Times previously disclosed that Nscale downplayed its relationship with ByteDance in its IPO materials — the latter contributed nearly three-quarters of its sales last year. Nscale publicly filed its prospectus on September 18, and the mandatory waiting period has long expired, but according to Bloomberg, the company is still "working through comments from the U.S. Securities and Exchange Commission" and has not yet launched a formal roadshow.
A Chain of Failures: New Cloud IPO Windows Closing One After Another
Nscale's predicament is not an isolated case but a microcosm of the ebbing tide in AI data center financing over the past month.
This wave of setbacks began on September 22: SoftBank's SB Energy, which had planned to raise funds for the world's largest data center, was forced to postpone its IPO due to investors balking at its valuation of over $50 billion, and it has still not launched formal marketing. Then came Firmus, which first cut its IPO price by 25% and then completely withdrew its listing plan; Philippine PLDT subsidiary Vitro postponed its data center REIT listing to 2027; and Blackstone's AirTrunk, even after raising its yield to approximately 6%, also delayed its Singapore REIT's $1.5 billion IPO to 2027.
The credit market had already sounded the alarm first. CoreWeave's credit default swap spreads have surged to a record high, implying a roughly 63% probability of default within five years; its 2032 unsecured bond yield has reached 15%. Gartner predicts that more than 80% of new cloud companies will cease independent operations before 2030. Goldman Sachs' credit team wrote in its weekend report:
"Capital has not disappeared; it has simply become more selective... The market is increasingly distinguishing between companies that can self-fund their growth and those whose growth depends on the continued availability of financing."
The Goldman Sachs high-yield credit team also noted that the average price of sampled high-yield data center project finance bonds has fallen from $101.5 in June to $94.7 in October, with "construction risk, refinancing exposure, tenant guarantees, power supply, and permitting" becoming core considerations in credit underwriting — and it is precisely power supply and permitting issues that directly caused Nscale to lose Microsoft as its anchor customer in Loughton.
Valuation Repricing Spreads to Equities: Tech Sector Hit by Successive Selloffs
Pressure from the credit market has fully transmitted to equity assets. CoreWeave lost approximately a quarter of its market value last week, while Nebius fell 32% over two weeks. According to a Goldman Sachs global equity derivatives team report, hyperscaler credit spreads have widened to their widest levels since 2025, and historically when spreads are in a similar range, related stock prices are approximately 15% lower than current levels.
Goldman Sachs technology, media, and telecommunications specialist Peter Callahan noted that the Nasdaq 100 recorded its worst weekly performance since April 2025 last week, falling approximately 550 basis points, with the semiconductor sector leading the decline at about 11.5%. The Philadelphia Semiconductor Index (SOX) has fallen for eight consecutive trading days, the longest losing streak since 2008; semiconductor sector valuations have compressed from approximately 29x forward earnings in the summer to about 17x. Goldman Sachs' AI beneficiary stock basket has retreated 25% relative to the broader market from its June high.
Next Step: Private Markets May Be the Only Way Out
For now, Nscale faces two paths. According to Bloomberg, discussions are ongoing, details of the IPO plan may change, and no final decision has been made.
One option is to accept a valuation discount and list at $16 billion to $19 billion — only slightly above the $14.6 billion valuation from its Series C funding in March this year. For Pre-IPO investors who injected a combined $3.36 billion in September, led by Dan Loeb's Third Point with participation from NVIDIA (NVDA), Apollo, and Citadel, this would be a far gloomier outcome than originally expected. The second option is to follow Firmus and turn to private markets — but even Blue Owl failed to find third-party funding for a $4 billion data center project.
The broader market backdrop also cannot be ignored. Bankers face a situation of "multiple intertwined risks," and the IPO market is freezing for nearly all AI projects except Anthropic. And Anthropic, as one of the two pillars of Nscale's order backlog, may itself face delays in its own listing plans. At that point, Nscale will be competing with one of its largest customers for the same pool of AI-themed investor capital — a rather delicate position.
When the pace of signing hundred-billion-dollar contract commitments far exceeds the pace of grid connections coming online, the three pillars supporting this wave of AI infrastructure construction mania — hyperscaler procurement commitments, cheap financing costs, and available power supply — are being repriced by the market one by one.
Comments