Anthropic, the maker of Claude, has signed a $10 billion, six-year compute procurement agreement with Volta Infra Holdings — a startup founded in January 2026 that has been in existence for roughly seven months — giving the frontier AI lab access to 121 IT megawatts of Nvidia Vera Rubin capacity at Bitdeer Technologies Group’s Tydal data center campus in Norway, according to Bloomberg. The financial mechanism that made the deal possible — a $1.3 billion credit backstop from JP Morgan affiliates and a second unnamed global institution — is not incidental to the story. It is the story: the same credit-decoupling function that Google’s $43.8 billion in data center lease guarantees performs for the TPU ecosystem has now arrived, through Volta, in the Nvidia ecosystem.
Bitdeer (Nasdaq: BTDR) disclosed the underlying deal terms Tuesday in a GlobeNewswire press release, announcing that its subsidiary Tydal Data Center AS has executed a 16-year colocation lease with Volta Tydal AS, a Volta subsidiary, worth approximately $4.7 billion in contracted base-term revenue. An optional eight-year renewal could bring the total potential contract value to approximately $8 billion over 24 years. Bitdeer shares surged as much as 23% on the news before settling closer to 10% above the prior close as the session progressed.
Volta: Infrastructure Finance as the Business Model
Volta Infra Holdings is not a data center operator. It does not build or own the physical facilities it deploys compute into — Bitdeer retains 100% ownership of the Tydal campus throughout the 16-year lease. Volta is not primarily a chip reseller. What Volta provides, in its own framing, is “the utility of compute”: the capital stack, credit support, and institutional financing relationships that allow a frontier AI lab to access purpose-built GPU capacity without needing to arrange infrastructure financing on its own balance sheet. Volta calls itself a fully integrated compute platform.
Ricard Boada, Volta’s CEO and a former infrastructure executive at Brookfield Asset Management, described the model in terms that deliberately echo the utility industry: “Compute has become a new infrastructure asset class, with AI models and applications as the verticals built on top. Every technology revolution has run on a physical layer beneath it. Railways carried industrialization.” His co-founder Sofia Gumuzio, also formerly of Brookfield, added that meeting AI demand “requires a platform that can mobilize infrastructure capital, secure power, and execute at industrial scale while moving at the pace of AI innovation.”
The mechanism Volta uses to deliver that utility is what distinguishes it from prior neocloud operators. Rather than requiring each AI lab customer to arrive with its own investment-grade credit rating — the condition that effectively limits direct data center leasing to hyperscalers — Volta assembles the credit support behind each deployment itself. In the Tydal arrangement, that means approximately $1.3 billion in standby letters of credit arranged by affiliates of JP Morgan and a second unnamed top-tier global financial institution, backstopping Volta’s payment obligations to Bitdeer.
This is the same credit-decoupling function that Google’s lease guarantee network performs in the TPU ecosystem — where Alphabet’s $43.8 billion in off-balance-sheet data center lease guarantees allow former crypto miners to issue investment-grade construction bonds, lowering their borrowing rates by a documented 2.2 percentage points relative to unguaranteed Nvidia-ecosystem operators. In Tydal’s case, JP Morgan’s letter-of-credit backstop substitutes for Alphabet’s guarantee, placing J.P. Morgan’s institutional creditworthiness between Bitdeer and any risk of Volta payment default. Bitdeer’s projected net operating income margin on the lease is approximately 90% — a figure reflecting the long-duration, contracted-revenue structure of the deal rather than any extraordinary operating efficiency.
How Vera Rubin Changes What Norway’s Hydro Power Is Worth
The hardware specified for the Tydal campus is Nvidia’s Vera Rubin platform — the company’s seventh-generation architecture detailed at GTC 2026, currently in production and entering partner deployments in the second half of 2026. A single Vera Rubin NVL72 rack packs 72 Rubin GPUs and 36 Vera CPUs into a fully liquid-cooled chassis, delivering 3.6 exaflops of NVFP4 inference performance and 2.5 exaflops of training compute. A full 40-rack Vera Rubin POD — the unit Nvidia uses for large-scale deployments — scales to 1,152 Rubin GPUs and 60 exaflops of aggregate compute. Dell Technologies is the hardware provider for the Tydal installation.
The Vera Rubin architecture matters for Norway specifically because of how it handles power. The NVL72 rack is fully liquid-cooled from the chip level — a requirement driven by the Rubin GPU’s transistor density fabricated on TSMC’s 3nm process node, which generates heat densities that air cooling cannot remove at scale. Norway’s cold climate and the Tydal campus’s hydroelectric power supply create conditions unusually well-matched to liquid-cooled AI hardware: ambient temperatures remain low for most of the year, reducing the energy needed to cool the liquid before it reaches the chips, and the hydroelectric grid delivers power without the emissions profile that has drawn regulatory scrutiny to natural gas-powered AI sites in other jurisdictions.
The Tydal facility is targeting a Power Usage Effectiveness ratio of approximately 1.1 — meaning that for every watt delivered to the Vera Rubin chips doing AI work, only 0.1 watts is consumed by cooling, lighting, and power conversion overhead. The average US data center runs at roughly 1.58 PUE; continental European facilities typically run between 1.3 and 1.5. Nordic facilities routinely achieve 1.1 to 1.2 because their cool ambient temperatures enable free-air or water-based cooling for most of the year, without mechanical chillers. At 121 IT megawatts of contracted load, the PUE gap alone between 1.1 and 1.5 represents roughly 48 megawatts of avoided overhead power — capacity that in Tydal goes back into powering additional compute rather than keeping heat exchangers running.
Norway generates over 90% of its electricity from hydroelectric sources, giving the Tydal campus a near-zero-carbon power profile without requiring renewable energy certificates or power purchase agreement offsets. Bitdeer has operated in Norway since 2018, acquired the Tydal site outright from its previous lease arrangement in 2024, and selected it specifically for its location within the Kirkvollen industrial site in Tydal municipality, Trøndelag county — an area where the grid supports large continuous loads. The campus’s excess heat is planned for use in food production on a neighboring property, an application of district-heating integration common at Nordic data centers.
Andreessen Horowitz Breaks Its Neocloud Rule for Volta
The $300 million Series A that Volta announced simultaneously with the Tydal deal carried investor names that signal how the infrastructure finance model landed in the venture community. Andreessen Horowitz and Altimeter Capital co-led the round, valuing Volta at $2.4 billion post-money. Nvidia and the family office of Michael Dell participated, with both also serving as strategic technology partners — meaning Nvidia supplied technology expertise and Dell’s commercial relationship provides hardware into Volta’s deployments.
Raghu Raghuram, managing partner at Andreessen Horowitz and a former CEO of VMware, was explicit about what Volta had to overcome to earn the firm’s participation. Andreessen Horowitz had previously avoided investments in so-called neoclouds and AI data center startups, but was drawn to Volta because of its founders’ experience in financing projects and securing power, Bloomberg reported, quoting Raghuram directly. Success, Raghuram said, would require “mastering the financial complexities of building and operating new clouds and selling to customers.”
The a16z blog post announcing the investment made the firm’s reasoning explicit: most large neoclouds had “followed the capital” and “sold capacity to hyperscalers” — acting as extensions of hyperscale infrastructure teams — while “companies without the hyperscalers’ balance sheets are left with the scraps.” Volta, the post argued, inverts that model by “assembling the credit support, project equity, and debt behind each deployment” rather than requiring each customer to provide its own institutional credit backstop.
That stated rationale maps directly onto the structural gap the Google guarantee network fills for Anthropic in the TPU ecosystem — where Alphabet’s off-balance-sheet commitments have effectively made it possible for former crypto miners to issue investment-grade bonds during construction, letting Anthropic access physical compute infrastructure without carrying that infrastructure debt. Volta’s model proposes to do the same in the Nvidia-chip supply chain, using institutional banking relationships rather than Alphabet’s balance sheet.
Altimeter Capital’s Jamin Ball, whose firm co-led the Series A alongside a16z, offered a less triumphant view of how many Volta-style entrants will survive the infrastructure cycle. “There’s going to be so many dead bodies and so much consolidation eventually,” Ball said, according to The Next Web. Volta is betting, Ball implied, that its banker relationships and blue-chip partners — JP Morgan, Dell, Nvidia, Andreessen Horowitz — are what keep it standing when demand slows.
Bitdeer’s Bitcoin-Zero Thesis, Validated
Bitdeer Technologies Group completed what may be its most consequential strategic move in February 2026, when it zeroed its Bitcoin treasury — selling every coin it owned and immediately selling all newly mined Bitcoin upon production, redirecting every dollar into data center infrastructure. That decision was controversial at the time: peers like MARA Holdings and MicroStrategy had doubled down on Bitcoin as a treasury reserve asset, while Bitdeer management insisted the sale was a liquidity-first pivot toward AI infrastructure opportunities.
Today’s announcement is the clearest validation yet of that thesis. The Tydal lease is Bitdeer’s first fully disclosed, fully conditioned major AI colocation deal — a 16-year contract averaging $202 per kilowatt per month across the contracted 121 IT megawatts, with 3% annual escalators built into both lease and services. Bitdeer estimates approximately $500 million in remaining capital expenditure to complete the build-out, at roughly $4 million per contracted IT megawatt — financing it plans to raise through additional debt, retaining 100% equity in the campus throughout.
The company is developing two additional data halls at the Tydal site, totaling 47 gross megawatts, for other AI and high-performance computing customers, with delivery targeted for the second half of 2027. It is also pursuing conversions at existing sites in Knoxville, Tennessee (37 MW) and Wenatchee, Washington (13 MW), with evaluations underway at its Rockdale, Texas facility (563 MW). The Tydal deal does not land Bitdeer inside Google’s credit guarantee network — the mechanism that has given TPU-ecosystem operators like TeraWulf a structural borrowing cost advantage — but it places the company inside a comparable institutional credit structure anchored by JP Morgan, in the Nvidia-chip supply chain where the Google guarantee does not reach.
Anthropic’s Compute Supply Chain Hits Seven Concurrent Paths
The Volta deal adds a seventh distinct compute relationship to Anthropic’s infrastructure portfolio — a level of supply-chain diversification that reflects both the company’s rapid revenue growth and its experience with capacity constraints that have forced usage limits on subscribers during peak demand periods.
Anthropic’s documented compute supply chain now spans: Amazon Web Services (Trainium2 chips, Project Rainier — more than 1 million chips); Google Cloud (Ironwood TPUs via a Compute SPV backed by Broadcom residual value guarantees and Apollo/Blackstone private credit, scaling from 1 gigawatt in 2026 to 3.5 gigawatts in 2027); SpaceX (Colossus 1 in Memphis, approximately $1.25 billion per month through May 2029, covering around 300 megawatts); AMD (Instinct MI450 Series GPUs via Helios rack-scale systems, up to 2 gigawatts beginning first half of 2027); Microsoft Azure (approximately $30 billion in committed spend); and now Volta/Bitdeer/Vera Rubin in Tydal, Norway. Reports indicate Anthropic is also in talks with Meta for a compute arrangement that could be worth as much as $10 billion over two years, though that deal has not been announced.
The interdependencies embedded in the Tydal structure are worth noting. Nvidia is simultaneously an investor in Volta and the chip supplier for the Tydal campus. Dell Technologies is simultaneously a Volta investor and the hardware provider for the site. Anthropic is the anchor revenue source underwriting the financial structure. These tangled relationships amplify returns when AI demand meets projections — and could amplify losses if it does not, in the same way the Google/Broadcom/TeraWulf network creates concentrated systemic exposure to Anthropic’s revenue trajectory.
Bitdeer itself carries approximately $1.9 billion in borrowings against first-quarter revenue of $188.9 million and a first-quarter net loss of $159.5 million. The company holds about $297.7 million in cash and restricted cash. The projected NOI margin of 90% — if the lease performs — would represent a substantial improvement in the underlying economics, but Phase I delivery is targeted for December 31, 2026, and Phase II for March 31, 2027, leaving only four months for Bitdeer to complete the first phase of a purpose-built Vera Rubin facility from its current construction state.
The June 29, 2026 preliminary announcement noted that the lease remained subject to “conditions precedent beyond Bitdeer’s control.” Tuesday’s full disclosure indicates those conditions have now been substantially satisfied, but Bitdeer’s own press release notes the transaction remains subject to closing conditions and is not yet effective. Volta’s Tydal subsidiary retains the right to terminate the contract without penalty after 10 years — a provision that materially reduces the minimum committed revenue from the announced $4.7 billion baseline.
Frequently Asked QuestionsWhat does Volta Infra Holdings actually do, and why does it need to exist?
Volta’s function is to assemble the credit support, project equity, and debt financing that allow an AI lab to access purpose-built GPU infrastructure without carrying that infrastructure debt on its own balance sheet. The problem it solves is that frontier AI labs — even companies at Anthropic’s scale, with $30 billion or more in annualized revenue — do not hold investment-grade credit ratings that would allow them to directly sign long-term data center leases or GPU procurement agreements on commercially viable terms. Volta steps between the lab and the infrastructure operator, arranging institutional backstops (in the Tydal case, $1.3 billion in JP Morgan letters of credit) that make the payment obligation creditworthy for the infrastructure owner. This is structurally identical to what Google’s $43.8 billion in lease guarantees do for Anthropic in the TPU ecosystem — except it operates in the Nvidia chip supply chain, where Google’s guarantee network has no footprint.
What is Nvidia’s Vera Rubin architecture and why is it significant for a facility this size?
Vera Rubin is Nvidia’s seventh-generation AI platform, announced at GTC 2026 and entering partner deployments in the second half of 2026. The core rack configuration, the NVL72, packs 72 Rubin GPUs and 36 Vera CPUs into a single fully liquid-cooled chassis, delivering 3.6 exaflops of inference compute and 2.5 exaflops of training compute. A full 40-rack Vera Rubin POD reaches 60 exaflops — exceeding the capacity of any prior publicly known supercomputer for AI-specific workloads. The architecture was built around the assumption that AI inference (“the age of inference” in Nvidia’s language) will dominate workloads as deployed models serve users rather than training new ones — making inference efficiency, not just peak training throughput, the governing design constraint. At 121 IT megawatts, Tydal would support a substantial cluster of Vera Rubin PODs, with Norway’s PUE 1.1 hydroelectric power environment allowing more of each megawatt to reach the chips rather than cooling overhead.
Why does Norway specifically make sense for frontier AI compute?
Norway generates over 90% of its electricity from hydroelectric sources — some of the cleanest and most reliably priced large-scale power available in Europe. Its cool ambient climate allows data centers to achieve Power Usage Effectiveness ratios of 1.1 to 1.2, compared with 1.3 to 1.5 in continental Europe and approximately 1.58 as the US average, meaning a larger fraction of every megawatt purchased actually powers AI chips rather than cooling equipment. For the Vera Rubin platform specifically, which requires liquid cooling at the chip level, Norway’s cool ambient temperatures reduce the energy cost of the cooling loop itself. The country also offers regulatory stability, EU data sovereignty compatibility for European AI workloads, and grid infrastructure built for large continuous industrial loads — exactly what AI GPU clusters require. As power has overtaken connectivity as the primary site-selection driver for data center development, Norway’s renewable, reliably priced, abundant hydroelectric capacity has drawn sustained attention from operators including Bitdeer, Polar Data Centres, OpenAI (Stargate Norway), and now Anthropic via Volta.
How does Anthropic’s Volta arrangement compare to its other compute deals?
The Volta/Bitdeer/Tydal arrangement is Anthropic’s first publicly confirmed dedicated AI compute deployment in Europe, and it uses a structurally distinct path: rather than leasing directly from a hyperscaler or signing a direct data center lease as the named tenant, Anthropic has contracted with a specialist infrastructure financier (Volta), which has in turn arranged institutional credit backstop from JP Morgan to support its own payment obligations to Bitdeer. This is a different architecture from the Compute SPV through which Google’s Ironwood TPUs reach Anthropic (where Broadcom provides residual value guarantees and Apollo/Blackstone provide private credit), but it serves the same function: decoupling Anthropic’s compute access from the company’s own creditworthiness. The result is a compute supply chain that now spans at least seven distinct paths — AWS Trainium, Google TPU, SpaceX Colossus, AMD Instinct, Microsoft Azure, and now Volta/Bitdeer/Vera Rubin in Norway, with Meta discussions reportedly ongoing.