AI data centers: which startup is ahead?

Which AI data center startups belong in this race?

Crusoe, Nscale, Fluidstack, and Lambda are the four private companies with the strongest claim to being full AI data center startups.

All four operate beyond basic GPU rental. They secure power or data center space, install dense computing systems, manage the software layer, and serve large AI customers. The comparison stays deliberately narrow: we want to know which startup can deliver a complete AI data center, not which company can offer access to a few hundred GPUs.

Nscale filed for a New York Stock Exchange listing in September 2026, and Lambda says it is aiming for a 2027 IPO. Until those listings happen, both remain private and stay in this ranking. CoreWeave and Nebius provide useful public-market benchmarks but do not belong here. Applied Digital, IREN, TeraWulf, Hut 8, and other former crypto miners are also public, and several of them now host capacity for these startups. Equinix, Digital Realty, Vantage, and QTS are established operators. Smaller GPU clouds such as TensorWave and Voltage Park remain credible, but their documented physical infrastructure is far below the scale of the four companies below. Established operators like Equinix and QTS run huge sites, and we look at the top companies today in our report covering data centers.

Funding needs a careful reading. Equity, convertible notes, GPU-backed loans, and project financing serve different purposes. We separate them wherever the disclosures allow it instead of adding every dollar into one flattering total.

Startup	What the company does	Completed or clearly disclosed funding	Additional infrastructure financing
Crusoe	Develops large AI campuses, builds modular data centers, secures energy, and operates Crusoe Cloud	A $3.9 billion Series F in September 2026 at a $30.9 billion post-money valuation, on top of roughly $3.9 billion of earlier equity and corporate debt	Large projects use separate structures, including a previously announced $15 billion Abilene joint venture
Nscale	Develops sovereign and hyperscale AI campuses while operating a managed AI cloud; filed for a NYSE IPO in September 2026	About $3.69 billion across its Series A, Series B, SAFE, and Series C rounds; valued at $14.6 billion in March 2026	$3.36 billion of convertible notes announced in September 2026, including a $1 billion Nvidia commitment due in November, on top of a $1.4 billion GPU-backed loan, $790 million of Norway financing, and a $900 million revolving facility
Lambda	Operates a mature GPU cloud and is moving into dedicated AI factories	More than $2.3 billion of equity after its Series E; reported in October 2026 to be raising up to $4 billion at a $14.5 billion pre-money valuation	A $1 billion senior secured credit facility, plus about $917 million reportedly borrowed against its GPUs
Fluidstack	Operates GPU clusters and develops custom AI infrastructure for frontier laboratories	An $830 million Series A at a $7.5 billion valuation (company-stated), followed by a reported $1.5 billion round led by Jane Street at about $18 billion in September 2026	No complete company breakdown of corporate debt and project financing is publicly available

Which AI data center startup is ahead today?

Crusoe is ahead in AI data centers today, and the lead has grown.

Crusoe told investors in September 2026 that it has more than 6 gigawatts of contracted capacity, about 1 gigawatt already operating, and more than $140 billion of total contracted value. These are company-stated figures. Even with that caveat, no private rival comes close to a gigawatt of live AI capacity.

Nscale is still the closest challenger on paper. Its IPO filing lists more than $103 billion of contracted value, with Microsoft and Anthropic accounting for about 85% of it. But Nscale had only about 55 megawatts running at the end of August 2026, and most of its contracts will not produce revenue for years. A backlog that is mostly paper is why we separate real customers from promises in our report on data centers.

Fluidstack has closed much of the gap with Nscale. It has Anthropic and Meta as customers, a reported $18 billion valuation, and reportedly nearly 1 gigawatt of power secured at converted bitcoin-mining sites. Lambda has the most mature conventional GPU cloud and a fast-growing backlog. Neither matches Crusoe on live megawatts, power development, and customer spread.

The structure is clear rather than a four-way tie. Crusoe leads overall, Nscale and Fluidstack fight for second place, and Lambda leads a narrower part of the market.

Who has the most AI data center capacity running now?

Crusoe has the clearest lead in AI data center capacity that customers can use today, by a wide margin.

The first two Abilene buildings for Oracle and OpenAI provide more than 200 megawatts of IT capacity, and Crusoe says a second phase of six buildings should take the campus to 1.2 gigawatts by the end of 2026. Crusoe now puts its total operating capacity at about 1 gigawatt of gross capacity. That figure comes from the company and has not been broken down site by site, so we treat it as company-stated.

Nscale’s numbers look very different once the IPO filing is read. At the end of August 2026, Nscale had about 7 megawatts active in its own data centers and 48 megawatts in rented facilities. Its Norwegian flagship at Glomfjord was running 6.4 megawatts of a planned 30, about six months behind schedule, according to reporting on the filing. Nscale had earlier described 30 megawatts there as operating; the filing-based figure is lower, and we rely on it.

Fluidstack says it manages more than 100,000 GPUs, many in third-party data centers. GPU management shows commercial reach, but it tells us little about how much power Fluidstack has developed itself. A reported investor memo projects 1.3 gigawatts of compute across more than 10 sites in 2026; that is a forecast, not delivered capacity.

Lambda operates production clusters across several locations but still does not publish one consolidated figure for live megawatts. Its announced sites with EdgeConneX in Chicago and Atlanta and its Kansas City facility cannot be counted as operating without a confirmed launch.

Startup	Best disclosed operating evidence	Comparable live-capacity index	What remains unclear
Crusoe	About 1 GW operating (company-stated, September 2026), including more than 200 MW at the first Abilene phase	100	The site-by-site split of the 1 GW and the pace of the Abilene second phase
Nscale	About 55 MW active at the end of August 2026 (7 MW owned, 48 MW rented), per its IPO filing	About 5	When Glomfjord reaches 30 MW and when the Texas and West Virginia sites go live
Fluidstack	More than 100,000 GPUs under management	Not comparable	Owned power, live campus capacity, and the split between direct and partner infrastructure
Lambda	Several production cloud regions, without a consolidated MW total	Not comparable	Total energized power across owned and partner sites

Who has the biggest AI data center pipeline?

Crusoe leads on contracted power, while Nscale has the biggest disclosed contract book among the challengers.

Crusoe reported more than 6 gigawatts of contracted capacity in September 2026. Named deals include about 1.6 gigawatts for Meta in Texas and Missouri, signed in June 2026, about 1.2 gigawatts for Oracle and 900 megawatts for Microsoft at Abilene. In September 2026, Jane Street reportedly signed a five-year cloud contract worth about $13 billion, Crusoe’s largest cloud deal so far. One project went the other way: in June 2026 Crusoe paused work on a 1.8-gigawatt Wyoming campus, and another operator is expected to take it over.

Nscale’s IPO filing lists more than $103 billion of contracted value. The two pillars are a $44.6 billion Anthropic agreement for compute at Nscale’s planned Monarch campus in West Virginia, signed in August 2026, and about $43.8 billion of Microsoft compute supply through 2033. The filing says the Anthropic deal depends on Nscale securing financing, and Anthropic can cancel if Nscale misses strict milestones. The Information reported that the contract figures shown to investors were called illustrative, not revenue guidance.

Fluidstack’s pipeline rests on Anthropic’s $50 billion American infrastructure program and newer work for Meta. Public disclosures still do not show how much of that spending becomes Fluidstack revenue or how many megawatts Fluidstack must deliver itself. In September 2026 it broke ground in Cameron County, Texas, on a $4 billion first phase planned for up to 1.5 gigawatts; that is planned capacity, not delivered capacity.

Lambda’s backlog of unfilled orders reportedly grew from $15 billion in June to about $50 billion in September 2026, according to the Wall Street Journal. That is a big jump, but Lambda has not said how much capacity or which customers sit behind it.

Startup	Strongest pipeline evidence	How we interpret it
Crusoe	More than 6 GW contracted and more than $140 billion of contracted value (company-stated, September 2026)	The strongest disclosed measure of customer-backed power capacity, spread across Oracle, Microsoft, Meta, and Jane Street
Nscale	More than $103 billion of contracted value in its IPO filing, including $44.6 billion from Anthropic	Very large on paper, heavily concentrated, and mostly years away from revenue
Fluidstack	Anthropic’s $50 billion American infrastructure program, plus Meta work	Exceptional upside, with too few disclosed commercial and capacity details
Lambda	Reported backlog of about $50 billion in September 2026	Fast growth, but capacity, customers, and timing remain undisclosed

Which startup has the strongest AI data center customers?

Crusoe has the strongest customer book because several large buyers have trusted it with separate, large-scale projects.

The operating Abilene phase connects Crusoe with Oracle and OpenAI. Microsoft selected Crusoe for a separate 900-megawatt Abilene campus with its own power plant, Meta signed for about 1.6 gigawatts in June 2026, and Jane Street reportedly committed about $13 billion to Crusoe Cloud in September. Later that month Crusoe announced multi-year deals with Perplexity and Thinking Machines Lab, without amounts. Four different large buyers is real diversification.

Nscale’s customers are excellent but concentrated. Microsoft and Anthropic account for about 85% of its contract book, and one unnamed customer made up 52% of first-half 2026 revenue. Microsoft’s exit from the Loughton site in October 2026 shows the cost of that concentration, even though Nscale says the relationship remains strong and its other contracts are unaffected.

Fluidstack’s selection by Anthropic for custom facilities in Texas and New York remains its anchor. It has added Meta and keeps serving Mistral AI, Poolside, and other AI developers. A second large infrastructure customer reduces the concentration risk that weighed on Fluidstack earlier in 2026.

Lambda has worked with Microsoft for years and has a multibillion-dollar infrastructure agreement with the company. It also serves a wide base of researchers, startups, enterprises, and AI laboratories. Crusoe still ranks first because its largest relationships already span an operating campus and several new developments.

Is Nscale’s $103 billion backlog as solid as it looks?

No. The contracts are signed, but Nscale’s IPO filing shows that most of that value is still a promise.

The growth is real. Nscale’s revenue reached $140.6 million in the first half of 2026, up from $10.4 million a year earlier, and its contract book went from about $100 million to more than $103 billion in roughly two and a half years. Few companies of any kind have signed that much demand that fast.

The costs are real too. Nscale lost $1.02 billion in the first half of 2026, almost three times the loss a year earlier. Most of its infrastructure is rented, its flagship Norwegian site is late, and the Anthropic contract can be cancelled if Nscale misses milestones or fails to raise the money to build. The UK setback in October shows that a signed customer does not guarantee a grid connection. A signed customer without a grid connection is a warning sign, and we follow who is falling behind in our data center market report.

Crusoe’s contract book carries less delivery risk: about 1 gigawatt is already operating behind its more than 6 gigawatts contracted, by its own count. Fluidstack moved from partner clusters to Anthropic and Meta projects, and Lambda’s reported backlog more than tripled between June and September. Nscale wins the speed contest on paper; Crusoe wins it in buildings.

Who can build and power AI data centers fastest?

Crusoe has the best evidence that it can build and energize a large AI data center quickly.

Crusoe built and energized two 100-megawatt buildings at Abilene in under one year. That required coordinating the shell, electrical systems, cooling, networking, substations, backup power, and GPU installation, not just placing servers inside an available colocation hall. The company now says it has about 1 gigawatt operating across its sites.

Crusoe has also built power into its operating model. Its agreements include roughly 750 megawatts of Bergen Engines generation, 12 gigawatt-hours of Form Energy storage, a 900-megawatt on-site plant with batteries for the Microsoft campus, and battery systems using Redwood Materials technology. It is not immune to setbacks: the Wyoming campus stalled in June 2026 after anchor-customer talks failed.

Nscale’s power strategy is strongest in northern Europe, where Glomfjord runs on hydropower and the 230-megawatt Narvik campus is planned around renewable supply. The record so far is weak: Glomfjord is behind schedule, and Loughton cannot get grid power until the 2030s. Nscale’s planned West Virginia campus uses an on-site microgrid, which may help, but it still has to be built.

Fluidstack says its integrated system can deliver gigawatts in six months, against an industry timeline of 18 to 24 months. We treat that as a company target. Its shortcut is partnering with converted bitcoin-mining sites, such as TeraWulf, Cipher Mining, and Hut 8, which already have power. That can be fast, but it is a different job from building a full campus from scratch.

Which startup has the most mature AI data center product?

Lambda has the most mature AI cloud product, while Crusoe has the most mature product for customers that need the whole data center delivered.

Lambda gives customers several clear ways to buy computing capacity. A small team can launch on-demand GPUs, a growing company can reserve a cluster, and a frontier laboratory can commission a single-tenant supercluster. Lambda advertises production-ready clusters from 16 to more than 2,000 GPUs and dedicated architectures that can scale from 4,000 to more than 165,000 GPUs.

Crusoe covers more of the physical stack. The company can find or generate power, develop the site, manufacture modular infrastructure, install the GPU systems, and run the workload through Crusoe Cloud. The reported Jane Street contract shows that the cloud layer, not only the buildings, now wins large deals. Crusoe has also added managed inference, serverless fine-tuning, and self-serve deployments.

Fluidstack’s product strength lies in managing large, dedicated clusters across infrastructure owned by different partners. Its managed Kubernetes and Slurm services include cluster-health monitoring and bare-metal orchestration. Nscale also offers GPU nodes, inference, fine-tuning, and sovereign AI factories, but its largest configurations remain tied to campuses scheduled for later delivery.

For cloud access today, Lambda is the easiest winner. For a customer asking one company to deliver the buildings, power, hardware, and cloud layer, Crusoe is further ahead.

Which AI data center startup performs best?

Fluidstack has the strongest independent performance evidence, although no startup has proved a universal speed advantage.

SemiAnalysis gave Fluidstack a Gold rating in its ClusterMAX 2.0 review after evaluating the company’s operating platform. The review covered around 80 providers and described Fluidstack as especially good at turning mixed third-party data center infrastructure into a dependable customer experience. That matters because networking, cluster health, scheduling, and support often decide whether expensive GPUs stay productive.

Crusoe’s strongest evidence comes from running high-density GB200 systems at Abilene with direct-to-chip liquid cooling, and from winning a demanding customer like Jane Street. In October 2026 it said its modular Spark units reached Nvidia Exemplar-level training results on B300 systems, a company-run test rather than a cross-vendor benchmark. Lambda provides immediately accessible B200 and H100 clusters and supports large single-tenant systems. Nscale has secured large allocations of GB300 and Rubin hardware, but its IPO filing counts about 25,000 GPUs deployed against roughly 461,000 active and contracted, so most of that specification sheet is still in the future.

These examples measure different things. SemiAnalysis tested operational quality, Crusoe demonstrated facility density, Lambda shows product availability, and Nscale emphasizes incoming hardware. There is still no neutral benchmark comparing training throughput, failure rates, network congestion, utilization, and cost across equivalent clusters.

Fluidstack wins the narrow performance question. That result alone is not enough to crown it the overall data center leader.

Which AI data center startup offers the best prices?

Lambda offers the clearest value for customers buying standard GPU cloud capacity.

Lambda publishes prices without requiring a sales call. Its listed on-demand rates start at $3.99 per GPU-hour for an eight-GPU H100 system and $6.69 for an eight-GPU B200 system. Smaller configurations cost slightly more per GPU, and Lambda does not charge data-egress fees. A customer can estimate a real budget before starting procurement.

Crusoe lists on-demand, reserved, and spot options for Nvidia and AMD systems, but much of the exact pricing still goes through sales. Crusoe has advertised spot capacity at discounts of up to 90% against hyperscaler on-demand rates. That can suit interruptible jobs, although a spot discount should never be compared directly with guaranteed reserved capacity. Spot capacity sold at steep discounts is one way to make money, and we weigh which business models are winning in our data centers report.

Nscale and Fluidstack negotiate their largest deployments privately. Nscale’s IPO filing gives total contract values but not per-GPU prices, and its first-half 2026 loss of $1.02 billion says nothing yet about long-run unit economics. We cannot judge their pricing without contract prices, utilization guarantees, power costs, and financing terms.

The answer changes for a custom campus. Crusoe may deliver the lowest total cost through integrated power and construction, but no startup publishes enough customer-level information to prove it. Lambda wins here because its economics are visible, comparable, and usable right away.

Which startup can scale AI data centers across the most countries?

Nscale has the broadest international AI data center pipeline, although Crusoe has delivered far more at hyperscale.

Nscale says it operates in 14 regions with a power pipeline of more than 10 gigawatts. It runs five facilities in Ireland, Norway, and Portugal, has eleven more in development, and has projects in Texas, West Virginia, and North Carolina. The European portfolio matters because governments and large companies want computing capacity and data to stay inside their own jurisdictions. The UK leg is weaker after Microsoft left Loughton in October 2026.

Crusoe remains concentrated in the United States, with large sites in Texas and Missouri and links to other states. Its Wyoming campus stalled in June 2026. Crusoe also runs cloud capacity in Iceland and uses modular systems for smaller sovereign and edge deployments.

Fluidstack manages clusters through third-party facilities in several regions, including Nordic deployments with Borealis, and is building in the United States with former bitcoin miners. That partner-led approach can expand quickly because Fluidstack does not need to own every building. Lambda follows a similar route through companies such as EdgeConneX and Cologix.

Nscale wins on geographic coverage. Crusoe remains ahead in proving that one company can repeat a very large physical development model.

What can Crusoe do that rivals cannot easily copy?

Crusoe has the strongest moat because competitors would need to copy several difficult businesses at the same time.

A conventional GPU cloud can buy Nvidia systems and lease data center space. A traditional developer can build a powered shell. Crusoe combines energy procurement, on-site generation, campus development, modular manufacturing, liquid cooling, hardware integration, and cloud software. Few startups control that whole route from electricity to a running AI workload.

Crusoe’s energy background matters more every month. Power availability has become the main limit on AI expansion, and Nscale’s Loughton site is a live example of a project with a customer but no power. Crusoe spent years placing computing near underused or stranded energy and now applies that to natural gas, grid connections, batteries, and wind-rich regions.

Nscale’s moat comes from sovereign positioning, access to capital, Nvidia’s backing, and its Microsoft and Anthropic contracts. Fluidstack’s operating software makes inconsistent partner infrastructure behave like one reliable cloud. Lambda benefits from a mature developer product, transparent pricing, and more than a decade of GPU experience.

Those advantages attract customers. Crusoe’s advantage reaches further into the physical bottlenecks that decide whether a project gets built at all.

Who has the strongest financing: Crusoe, Nscale or Fluidstack?

Crusoe has the strongest financing position, because its capital is backed by live capacity. Nscale has raised a lot, but more of it is debt and more of its spending still lies ahead.

Crusoe raised $3.9 billion in the first close of its Series F in September 2026 at a $30.9 billion post-money valuation, co-led by Atreides Management, Mubadala Capital, and Valor Equity Partners, with Nvidia, Founders Fund, and QIA also investing. It combines that equity with corporate credit, project debt, and joint ventures, so outside capital can finance individual campuses. The operating Abilene phase gives lenders a completed project to judge future ones against. With equity, corporate credit and project debt all in the mix, we track who is getting funded, and by whom, in our report on data centers.

Nscale has about $3.69 billion of equity and SAFE financing, more than $3 billion of earlier loans and credit lines, and $3.36 billion of convertible notes announced in September 2026: $2.36 billion closed at once, and $1 billion is an Nvidia commitment due in mid-November. Nvidia has also agreed to guarantee $860 million of Nscale’s Texas lease obligations. Its NYSE IPO aims to raise about $3 billion. Reports differ on the target valuation: the Financial Times said about $35 billion, while CNBC said about $30 billion.

Fluidstack’s reported $1.5 billion round led by Jane Street in September 2026 values it at about $18 billion, up from $7.5 billion earlier in the year. Reports on the earlier round differ: Fluidstack itself cites $830 million, while some press reports said $750 million. The Wall Street Journal also reported talks on a $5 billion Pentagon loan to Fluidstack; nothing has been approved. Lambda’s $2.3 billion-plus equity, $1 billion credit facility, and reported $4 billion pre-IPO raise fit a company expanding from cloud into larger dedicated sites.

Crusoe gets the edge because its capital has already produced a large working asset. Nscale may soon have more cash from public markets, but its spending program remains several steps ahead of its completed infrastructure.

What could knock Crusoe out of first place?

Nscale can overtake Crusoe only if its Anthropic and Microsoft deployments arrive on schedule, while Fluidstack needs to prove its Anthropic and Meta buildouts at scale.

Nscale’s route to first place is clear but steep. It must energize the Texas, West Virginia, Norway, and Portugal projects, meet the Anthropic milestones, and replace the lost Loughton tenant. Nscale says it is looking for other power sources for Loughton.

Fluidstack faces a different problem. Anthropic and Meta give it extraordinary demand, but it relies heavily on partner sites and discloses little about completed megawatts or project economics. Its reported revenue projections remain far below its valuation, and New York’s pause on new hyperscale data center permits remains a possible risk for its New York project; public information does not show whether that site is affected.

Crusoe still carries serious risk. Its contracted portfolio needs several campuses and power systems to move at the same time. It relies on gas generation for some developments, which brings environmental opposition, fuel-price exposure, and permitting disputes. The Wyoming pause and Microsoft’s replacement of an abandoned OpenAI expansion at Abilene show that big customers can move capacity elsewhere.

Lambda’s main threat is competition. GPU clouds are crowded, and Microsoft, Google, Amazon, Oracle, CoreWeave, Nebius, and newer providers all pressure pricing. Lambda’s mature cloud and growing backlog make it resilient, but the physical data center ranking will stay hard to win without much larger disclosed campuses.

Which AI data center startups are actually ahead?

Crusoe is the clear AI data center leader, Nscale holds second place by a narrowing margin, Fluidstack is close behind, and Lambda wins a narrower part of the market.

Crusoe takes first place because it has done the hardest part at scale: about 1 gigawatt operating, by its own count, for serious customers. Around that base it has added more than 6 gigawatts of contracted power, new customers such as Meta and Jane Street, a $30.9 billion valuation, dedicated generation, and cloud software. The lead rests on buildings and customers, not on fundraising.

Nscale keeps second place on the strength of its contract book: more than $103 billion, with Anthropic and Microsoft behind most of it. But its IPO filing showed only about 55 megawatts active, a late Norwegian flagship, and a $1.02 billion first-half loss, and Microsoft has since left the Loughton site. Delivery will decide whether it stays there.

Fluidstack ranks third, close to Nscale. Anthropic and Meta give it two anchor customers, a reported $18 billion valuation gives it capital, and independent testing supports its operating platform. It would move up once it discloses completed campus capacity and firmer project economics.

Lambda ranks fourth in this comparison, even though it offers the strongest standard cloud-buying experience. It has real customers, transparent pricing, a reported $50 billion backlog, and a planned 2027 IPO. Crusoe, Nscale, and Fluidstack are simply making bigger bets on complete physical campuses.

Rank	Startup	Why the startup holds this position	What could change the rank
1	Crusoe	About 1 GW operating and more than 6 GW contracted (company-stated), a $30.9 billion valuation, power development, and a diversified customer book	Delays across its campus portfolio, gas-power opposition, or more stalled projects like Wyoming
2	Nscale	More than $103 billion of contracted value, including Anthropic and Microsoft, international reach, and Nvidia backing	Late sites, cancelled Anthropic milestones, or another lost tenant could drop it below Fluidstack; fast delivery could move it closer to Crusoe
3	Fluidstack	Anthropic and Meta as anchor customers, a reported $18 billion valuation, and the best independent cloud-operations rating among the four	Disclosed, completed campus capacity could push Fluidstack above Nscale
4	Lambda	The most mature GPU cloud, the clearest pricing, and a reported $50 billion backlog	Lambda needs to disclose and deliver a much larger physical data center footprint