← All posts / Industry

Three-Quarters of Its Revenue, One Buried Filing: How ByteDance Used Nscale's Norway Cloud to Reach Nvidia's B200s

An FT investigation reveals ByteDance accounted for ~75% of Nscale's 2025 sales by renting B200 GPUs in Norway — a US export-control loophole disclosed only in a sidecar IPO filing.

Three-Quarters of Its Revenue, One Buried Filing: How ByteDance Used Nscale's Norway Cloud to Reach Nvidia's B200s

Just days before one of the most anticipated AI infrastructure listings of the year, the Financial Times has published an investigation that turns Nscale’s IPO pitch inside out. The British AI cloud provider — Nvidia-backed, targeting a valuation of roughly $30–35 billion on the New York Stock Exchange under the ticker NSCL — spent its main prospectus emphasizing blue-chip Western customers like Microsoft and Anthropic. But according to the FT, the customer that actually built Nscale’s business was Beijing-based ByteDance, which accounted for nearly three-quarters of the startup’s 2025 sales and was quietly renting advanced Nvidia GPUs through Nscale’s Norwegian facilities — chips it could never have legally bought in China.

The revelation, drawn from SEC filings published Friday as part of Nscale’s IPO documentation, lands at a uniquely awkward moment. Nscale is courting public-market investors with a story about disciplined, infrastructure-first growth. Its founder, 32-year-old Josh Payne, wrote in his letter to prospective investors that the company was “building against contracted customer demand, underwriting projects to attractive long-term returns, maintaining prudent leverage.” What the main pitch didn’t foreground was how much of that early contracted demand came from a Chinese tech giant threading its way through a gap in US export controls.

The mechanics of the loophole

The core of the FT report is contractual, not speculative. SEC filings show that in May 2025, Nscale signed a contract with an entity called Spring, which agreed to rent 2,304 of Nvidia’s advanced B200 chips from the UK-based cloud provider. Spring, per the FT’s reporting, ties back to ByteDance. Because the GPUs physically sat in Nscale’s Norwegian data centers — outside both China and the US — ByteDance could access top-tier Nvidia compute without ever touching a restricted export. Renting remote compute is not the same as importing chips, and that distinction is precisely where the strategy lives.

Norway matters here for reasons beyond geography. Nscale operates Arctic Circle data centers in Glomfjord and elsewhere in the country, powered entirely by renewable hydroelectric energy — a fact the company markets heavily. Cheap, green power and cool climates made Norway the foundation of Nscale’s early capacity. It also, apparently unintentionally, made it the perfect jurisdiction for a Chinese customer that US rules were designed to exclude from frontier AI silicon.

This is not ByteDance’s only workaround. Earlier this year, the Wall Street Journal reported the company was working with a Malaysian cloud provider to acquire 36,000 B200 chips for AI R&D — an estimated $2.5 billion arrangement — and reports have circulated of a planned ~$14 billion Nvidia chip budget for 2026 alongside a 1,000-person chip team relocated to Singapore under the Picoheart banner. The pattern is consistent: where direct purchase is blocked, assemble compute abroad. What makes the Nscale case different is scale relative to the supplier — ByteDance wasn’t one customer among many; for a formative year, it was the business.

What the S-1 actually says

Nscale filed publicly on September 18, 2026. The headline numbers that carried coverage were spectacular: revenue of $140.6 million in the first half of 2026, up 1,252% from $10.4 million a year earlier; a net loss of $1.02 billion; more than $103 billion in total contracted value across 461,000 active-or-contracted GPUs; 14 operating regions and a power pipeline exceeding 10 gigawatts. Reuters noted the company separately agreed to sell $3.1 billion of convertible bonds, including $1 billion to Nvidia itself, and that Nvidia guaranteed up to $860 million of Nscale’s obligations on a Texas data center lease.

The prospectus also contained a concentration warning: a single unnamed customer accounted for 52% of revenue in the first half of 2026, with Microsoft and Anthropic expected to become major customers going forward. The ByteDance connection, however, was not part of the main narrative. Per the FT, the ties were disclosed only in a sidecar filing published alongside — technically available, practically invisible. Nscale also discloses over $2 billion in Nvidia commitments and acknowledges the risk that tighter US export controls could disrupt its business, a risk factor that reads very differently once you know who was paying the bills.

Why this matters beyond one IPO

First, it exposes a structural blind spot in export-control architecture. US rules restrict the sale of advanced chips to Chinese entities; they are far weaker on Chinese entities renting those chips inside third countries. If a company like ByteDance can quietly assemble tens of thousands of GPU-hours of frontier compute through European neoclouds, the control regime governs paperwork, not capability. Expect regulatory attention: the FT report is exactly the kind of case study that motivates Commerce Department revisions, and Traders Union notes the export-control risk factor now looks less like boilerplate and more like a live threat to Nscale’s revenue history.

Second, it complicates the AI-infrastructure investment story at a sensitive time. Neoclouds like Nscale, CoreWeave, Nebius, and Lambda are competing for capital partly on the quality of their customer books. Renaissance Capital strategist Matt Kennedy told Reuters the setup for AI infrastructure deals is “good enough to get these deals done, but it’s nothing like the euphoria of a few months ago.” A revelation that ~75% of 2025 revenue came from a Chinese giant exploiting a trade loophole — weeks before the bell rings on NYSE — is the kind of due-diligence surprise that reprices trust, even if the customer mix has since shifted toward Microsoft’s 200,000-GB300 GPU commitments and Anthropic’s $45 billion West Virginia deal.

Third, it puts Nvidia in an uncomfortable position. Nvidia is Nscale’s investor, creditor ($1 billion in convertible bonds), and lease guarantor — and its chips were the product flowing through the loophole. Nvidia has publicly committed to complying with US export rules; having its own portfolio company serve as the channel is a reputational exposure it will have to address, likely with new know-your-customer diligence on neocloud tenants.

The road ahead

Nscale’s trajectory from cryptocurrency-mining spinout (it emerged from Arkon Energy in 2024) to $30-billion-plus IPO candidate in two and a half years is genuinely remarkable, and its contracted backlog is real. Anthropic’s $45 billion commitment, Microsoft’s multi-country GPU rollout, the pending Anyscale acquisition, and a board featuring Fidji Simo, Nick Clegg, and Sheryl Sandberg give it a credible Western anchor-story. But the FT’s reporting recontextualizes the foundation: the early revenue that proved the model came substantially from the very kind of customer US policy is designed to fence off — reachable only through a jurisdictional gap that regulators are now certain to scrutinize.

For investors reading the S-1, the lesson is the sidecar filing. For policymakers, the lesson is that chip controls which stop at ownership boundaries don’t stop compute access. And for the broader industry, it’s a reminder that in the AI buildout, the most important disclosures are sometimes the ones a company is technically required to make — and structurally motivated to bury.