AI cloud: which startup is ahead?

Which startups actually count as AI cloud companies?

Eight private startups belong in the serious AI cloud comparison today, spread across three distinct parts of the market.

We include startups that sell direct access to GPUs, managed training, production inference, or purpose-built AI data centers. Fireworks AI, Together AI, and Baseten sit closest to the model-serving layer. Lambda and Runpod sell more direct access to compute. Crusoe, Nscale, and Fluidstack combine cloud services with large infrastructure projects. With players split between model serving, raw compute and big builds, we name the companies that lead today in our report on AI infrastructure.

We exclude AWS, Microsoft Azure, Google Cloud, Oracle Cloud, and other established hyperscalers. CoreWeave and Nebius also sit outside the ranking because both are public companies. They remain useful benchmarks. CoreWeave generated $5.13 billion of revenue in 2025 and ended that year with a $66.8 billion backlog, according to its annual results. Nscale filed for a US IPO in September 2026 but has not listed, so it stays in the field for now. Lambda has reportedly told investors it plans to go public in 2027.

Funding needs a warning label. Software-heavy providers mostly raise equity, whereas data-center builders also use equipment loans, project finance, convertible notes, and customer-backed debt. The figures below show the approximate disclosed capital we could verify and flag the companies where debt makes the comparison messy.

Startup	What it does	Approximate disclosed funding	Main lane
Fireworks AI	Custom model training and production inference	About $1.83B in equity, including a $1.505B Series D in July 2026	Inference cloud
Together AI	Inference, fine-tuning, training, and GPU clusters	About $1.33B in equity, including an $800M Series C in July 2026	Full-stack model cloud
Baseten	Deploys and operates custom models across multiple clouds	More than $2B in equity	Inference platform
Lambda	On-demand GPUs, managed clusters, and private supercomputers	About $2.3B in equity, plus debt including a $1B senior secured financing in October 2026; a round of up to $4B is reported in talks	GPU cloud
Nscale	Builds AI factories and sells cloud infrastructure	About $7B across equity, SAFE, and convertible rounds, plus several debt facilities; US IPO filed in September 2026	Hyperscale infrastructure
Crusoe	Develops power, data centers, GPU clusters, and cloud software	More than $7B across debt and equity, including a $3.9B Series F in September 2026	Vertically integrated AI cloud
Runpod	Self-service GPUs, serverless inference, and clusters	About $122M in equity	Developer cloud
Fluidstack	Builds and operates custom AI data centers	More than $2.6B in total funding, including a $1.5B round at an $18B valuation in September 2026 (reported); a $5B Pentagon loan is reported in talks	Custom infrastructure

Is there a clear AI cloud leader today?

Lambda is narrowly ahead overall today, Fireworks AI leads production inference, and Together AI is third.

The answer still changes as we move across the stack. Lambda now combines the most mature large-cluster product with the biggest reported contract book of any operating private cloud: a $35 billion Anthropic deal reported in late August 2026 and about $50 billion of unfilled orders in September, according to The Wall Street Journal. Bloomberg reported in August that its 2026 revenue was on track to top $1.5 billion.

Fireworks has the best company-stated numbers in inference: more than $1 billion in annualized revenue and more than 40 trillion daily tokens. Together is close on bookings and covers more of the training-to-inference workflow. The lead is narrow because Lambda’s key figures come from press reports, while Fireworks publishes its own.

Crusoe, Nscale, and Fluidstack lead a separate contest: building future capacity for hyperscalers and frontier labs. Nscale’s IPO filing shows how wide the gap between contracts and live capacity can be. We give operating workloads more weight than planned campuses when deciding who is ahead now.

Part of the market	Leader now	Closest challenger	How clear is the lead?
Production inference	Fireworks AI	Together AI, Baseten	Fireworks has the strongest company-stated revenue and token figures
General GPU cloud	Lambda	Crusoe, Runpod	Lambda pairs the most mature cluster product with a reported $50B order backlog
Future physical capacity	Nscale	Crusoe, Fluidstack	Nscale leads disclosed contracts, but its IPO filing shows about 55MW operating
Developer self-service	Runpod	Lambda	Runpod reports more than $325M in ARR and 1.2 million developers
Overall private AI cloud	Lambda	Fireworks AI, Together AI	A narrow lead built on reported figures; it could change with the next disclosure

Who is making real money from AI cloud services?

Lambda appears to make the most money in the private field, but Fireworks has the largest revenue figure a company has stated itself.

Bloomberg reported in August 2026 that Lambda’s revenue for the year was on track to top $1.5 billion. Lambda has not published that number, and we found no second independent report of it. Sacra’s last public estimate, from 2025, put it near $505 million of annualized revenue, so the reported figure implies roughly a tripling. Lambda sells capacity by the hour and under long contracts, which makes its revenue easier to forecast than usage-based inference, but also more tied to a few large buyers.

In July 2026, Fireworks said it had passed $1 billion in annualized revenue. Its previous major disclosure put the figure above $280 million, so the run rate grew at least 3.6 times between the two announcements. The company also says 95% of its tokens now come from models specialized for customer data and specific jobs. Those workloads sit inside real products, where switching providers is an engineering project.

Together AI operates in the same range, although its best public figure measures bookings. Annual bookings crossed $1.15 billion, and Sacra estimates roughly $1 billion in annualized revenue. Bookings and recognized revenue remain different measures. The HUMAIN forecast of more than $5 billion of gross annualized revenue applies to a data center that is not built yet.

Baseten was reported in June 2026 at a run rate near $600 million. Runpod said in October 2026 that it had passed $325 million in annual recurring revenue. Nscale’s IPO filing gives the first hard number from the builders: $140.6 million of revenue in the first half of 2026, up 1,252% year on year, against a net loss of about $1.02 billion. Crusoe and Fluidstack still publish no comparable revenue figure.

Which AI cloud startup is growing fastest now?

Lambda is growing fastest in contracted demand; Fireworks is growing fastest on company-stated revenue and usage.

Lambda’s reported backlog rose from about $15 billion in June 2026 to about $50 billion in September, mostly on the Anthropic agreement, according to a letter to investors seen by The Wall Street Journal. That is contracted future revenue, not revenue earned. It still shows that the largest buyers now treat Lambda as a primary supplier, not a spillover option.

Fireworks moved from more than $280 million to above the billion-dollar mark in annualized revenue. Daily usage climbed to more than 40 trillion tokens. Fireworks gave the earlier baseline as more than 10 trillion tokens, while a July 2026 report on the round cited 15 trillion. Either way, revenue and workload grew together, which is stronger evidence than a single giant contract.

Baseten is sprinting from a smaller base. In September 2026, CEO Tuhin Srivastava said token volume had grown about 40 times and revenue about 10 times over twelve months. The company had earlier described twentyfold revenue growth, so the two company statements do not match, probably because they cover different periods.

Runpod offers the clearest developer-growth pattern. It served more than 500,000 developers when it announced $120 million in recurring revenue in January 2026. In October 2026, it said recurring revenue had passed $325 million and that more than 1.2 million developers use the platform. It also said agents make up 8% of its users but 24% of its revenue.

Together is moving quickly too. In October 2026, it added a dedicated B300 inference cluster with IBM Cloud and NVIDIA and launched a post-training platform. Its test remains turning bookings and new capacity into recognized revenue.

What does Anthropic’s $35 billion deal with Lambda change?

It makes Lambda the private cloud with the largest contract from a frontier lab, and it makes Anthropic the most important buyer in this market.

According to Bloomberg and The Wall Street Journal, the agreement is worth about $35 billion over roughly six years and covers about 350 megawatts of Nvidia hardware at Hut 8’s campus in Nueces County, Texas. The Journal reported that Nvidia holds the lease on the site. Lambda already had a multibillion-dollar Microsoft agreement and, since May 2026, a multi-year supply deal with Hudson River Trading. The new contract moves it from strong GPU vendor to core supplier for one of the three leading model labs. Becoming core supplier to a top lab changes a company's standing; we compare who is ahead in our report on who's ahead in AI infrastructure.

Anthropic is now spreading its compute across several startups. Fluidstack was selected to build custom data centers in New York and Texas. Nscale says Anthropic agreed to take more than 400 megawatts at its West Virginia campus under a six-year deal. Crusoe, for its part, was reported in September 2026 to have signed a five-year cloud deal worth about $13 billion with Jane Street, and it announced partnerships with Perplexity and Thinking Machines Lab.

Fireworks, Together, and Baseten still hold the strongest set of AI-native production customers: Cursor, Harvey, Notion, Decagon, Abridge, and OpenEvidence among them. These logos matter because inference speed and cost sit inside the customers’ products.

The same companies remain customers, suppliers, investors, and rivals. Nvidia sells the chips, holds the lease behind Lambda’s Anthropic capacity, and invests in Lambda, Nscale, and others. Microsoft buys from Lambda and Nscale while selling Azure. A large contract is real demand, but it is not independent demand.

Can Nscale deliver the compute behind its IPO filing?

Not yet: Nscale has the largest disclosed future commitments, but its own filing shows only a small share of that capacity running today.

Nscale filed for a US IPO on September 18, 2026. Reports on the filing cite about 461,000 GPUs active or contracted, five data centers with about 55 megawatts operating, and 12 more sites in development. Microsoft has contracted 1.35 gigawatts at the West Virginia campus, where the first 2 gigawatts are expected online in the first half of 2028. Sources disagree on the contract total: Bloomberg put it near $51 billion in August, while reports on the filing cite $103 billion of contracted value.

The financing is just as large. In September 2026, Nscale raised $3.36 billion in convertible notes led by Third Point, including $1 billion from Nvidia due by mid-November. It is reported to be targeting a valuation of about $30 billion at listing. The gap between 55 megawatts and gigawatt-scale promises is the whole question for buyers of the stock and for customers waiting on capacity.

Crusoe measures scale in power and has more of it built. The company says it has contracted 4.9 gigawatts, with more than 40 gigawatts in its wider pipeline. Its first Abilene buildings are operating, and in July and September 2026 it announced new Texas campuses in Childress and Armstrong County. Crusoe also opened a second manufacturing plant in Tulsa in September 2026, which gives it more control over electrical equipment than a provider that leases finished space.

Lambda has the clearest purchasable product. Its self-service clusters range from 16 to more than 2,000 GPUs, and private superclusters run from 4,000 to more than 165,000 GPUs. In September 2026 it announced a new data center in Mayes County, Oklahoma. Runpod serves the opposite end through on-demand GPUs, serverless endpoints, and multi-node clusters. Fluidstack’s largest Anthropic facilities are still rolling out, and in September 2026 it broke ground on a $4 billion first phase of a Texas campus in Cameron County.

Which AI cloud is strongest for training large models?

Lambda leads large-model training among private AI clouds, and the Anthropic deal strengthens that lead.

Its product already spans single instances, self-service clusters, managed Slurm and Kubernetes, and private supercomputers with thousands of GPUs. In October 2026, it said it would build NVIDIA’s open-source Run:ai scheduler into its managed Kubernetes service. Agreements with Microsoft and now Anthropic confirm demand at the level of tens of thousands of GPUs. That availability and buyer trust put it ahead of startups whose largest systems remain construction projects.

Crusoe is the closest infrastructure challenger. It controls more of the physical chain: energy, campuses, electrical equipment, networking, cloud operations, and training software. Its September 2026 partnership with Perplexity covers the full model lifecycle, which shows it can win training work beyond hyperscale leases. A frontier lab that wants a dedicated campus and long-term power may prefer Crusoe’s model.

Among the inference-led companies, Together has the best training offer. It provides managed GPU clusters, pre-training, fine-tuning, reinforcement learning, and serving in one place, and it launched a dedicated post-training platform in October 2026. Fireworks made its training API generally available in September 2026, which narrows the gap at the fine-tuning end.

Nscale could take this position later. Its planned systems are larger and tied to Microsoft and Anthropic. Today, Lambda has far more evidence of a training cloud customers can use now.

Who is winning the AI inference cloud race?

Fireworks AI is the current inference leader, with Together close enough to make this a real two-company race.

Fireworks says it now serves more than 40 trillion tokens each day, and that 95% of them come from models adapted to customer data and tasks. Its work with Cursor and Harvey places that customization inside high-value products. In September 2026, it added multi-region deployments and published its own index of specialized models, a sign that it is selling routing and tuning, not just cheap tokens.

Together reported more than $1.15 billion in annual bookings and says it serves over one million developers. Decagon reported inference costs falling sixfold and voice latency dropping below 400 milliseconds on its stack. Cursor uses Together’s Blackwell infrastructure for production inference. The new IBM Cloud cluster adds dedicated enterprise capacity.

Baseten is a serious third player. Token volume grew about 40 times in a year, according to its CEO, and it names Cursor, Notion, Harvey, Abridge, and OpenEvidence as customers. It sources compute across more than 20 cloud providers, which gives customers portability, although Baseten owns less of the physical stack than Lambda or Crusoe.

Direct token comparisons remain imperfect because providers count input, output, cached, and reasoning tokens differently. We read workload figures alongside revenue, customers, and deployment depth. Taken together, the evidence puts Fireworks first, Together close behind, and Baseten clearly third.

Which AI cloud gives customers the best economics?

Runpod usually offers the cheapest route to raw GPUs, while Fireworks and Together create larger savings when inference software dominates the bill.

Lambda’s public pricing is a useful baseline. An eight-GPU H100 SXM instance has been listed at $3.99 per GPU-hour and a single H100 SXM at $4.29, with no egress fees. Runpod uses a broader and more variable supply base, so customers can often find lower hourly prices for flexible workloads.

Production inference has a different cost structure. Utilization, batching, quantization, caching, model choice, and engineering time can matter more than the sticker price of one GPU-hour. Together says Decagon cut inference costs sixfold. Fireworks publishes cost-per-task comparisons for new open models such as DeepSeek V4 Pro, and Baseten customers can mix open and closed models to cut costs where the quality gap is small. Cheaper inference and open-plus-closed model mixes are different ways to earn a living; we compare which business models win in our AI infrastructure report.

The claims are hard to standardize. Each provider picks favorable models, latency targets, and hardware. Buyers should test their own traffic and compare cost per completed task, p95 or p99 latency, failure rates, and engineering effort. For a small team buying flexible compute, Runpod has the clearest price advantage. For a company serving millions of model calls, Fireworks or Together can save more by squeezing extra work from each GPU.

Who is using AI cloud funding most efficiently?

Runpod has built the most business with the least equity, and Lambda is the most efficient of the large players.

Runpod reached $120 million in recurring revenue after raising about $22 million. By October 2026, it reported more than $325 million in annual recurring revenue on about $122 million of total equity, a ratio above 2.5. That is unusual for an infrastructure startup.

Lambda’s reported 2026 revenue of more than $1.5 billion against about $2.3 billion of equity gives a ratio near 0.65, better than Fireworks at around 0.55 on its stated run rate. Together’s outside revenue estimate against about $1.33 billion of equity produces a ratio near 0.75, though that revenue figure is unverified. Lambda leans more on debt than the inference platforms, including its $1 billion senior secured financing in October 2026.

Infrastructure builders need a separate test. We ask whether each dollar is backed by a signed customer, a financeable site, and a believable delivery schedule. Crusoe gives the strongest physical evidence. Nscale has the largest commitments, but its filing shows a $1.02 billion net loss in the first half of 2026 on $140.6 million of revenue. Fluidstack has one exceptional customer and far less diversification. A venture backer’s memo projects its 2026 revenue at $660 million, a forecast, not a reported figure.

What can the leading AI cloud startups do that competitors cannot easily copy?

Crusoe owns the hardest physical advantage to copy, Lambda has the deepest frontier-lab relationships among operating clouds, and Fireworks makes customers hardest to pull away.

Crusoe works across energy, site development, construction, electrical equipment, GPU infrastructure, and cloud software. Rebuilding that chain would take years of permits, power agreements, factories, and operating experience. Its $3.9 billion Series F and second Tulsa plant add capital and manufacturing depth to that system.

Lambda’s edge is trust at scale. Microsoft and Anthropic have both handed it multibillion-dollar workloads, and its close ties to Nvidia help it secure scarce hardware. Few rivals can show that record with frontier buyers.

Fireworks’ advantage sits closer to the customer. When a company fine-tunes a model, builds evaluations, connects proprietary data, and tunes production traffic on one platform, moving becomes a real engineering project. Specialized models account for 95% of its traffic.

Together combines systems research with training, post-training, clusters, and inference. Baseten hedges supply risk across more than 20 clouds. Nscale is building leverage through power access and giant commitments. Runpod’s developer base gives it distribution and a steady flow of smaller workloads. These advantages can still erode: hyperscalers have their own chips and larger balance sheets, open-source serving software keeps improving, and Nvidia influences who gets scarce hardware.

Where could the AI cloud leaders fail?

Customer concentration and construction risk could break the infrastructure leaders, while price compression is the bigger threat for inference clouds.

Concentration rose this year. Lambda’s two largest disclosed contracts come from Anthropic and Microsoft. Nscale’s contract book depends mostly on the same two buyers. Fluidstack’s rise is tied to Anthropic. Crusoe has more customers but still relies on a few hyperscale campuses. A delayed facility or a changed customer plan can push billions of expected revenue into another year.

The physical economics are unforgiving. New GPU generations arrive quickly, while the debt used to finance a cluster can last several years. A provider that buys too early, overpays for power, or misses its delivery window may rent older hardware at lower prices before the financing is repaid. CoreWeave is the public warning: it produced $5.13 billion of revenue in 2025, yet interest expense reached $1.23 billion and it reported a $1.17 billion net loss. Nscale’s first-half 2026 loss shows the same pattern at an earlier stage.

Inference providers face falling software prices. vLLM, SGLang, TensorRT-LLM, quantization, and hyperscaler discounts keep improving. Fireworks, Together, and Baseten need to create value through tuning, customization, reliability, and workflow integration. Plain access to a popular open model is becoming a commodity.

Nvidia’s many roles complicate the picture. It supplies chips, invests in Lambda and Nscale, committed $1 billion to Nscale’s latest notes, and reportedly holds the lease on the Texas site behind Lambda’s Anthropic deal. Those links help startups get hardware and capital, but they make demand look more independent than it is.

Is Lambda now the AI cloud startup with the most momentum?

Yes. Lambda has more momentum than any private AI cloud currently, with Fireworks and Crusoe close behind.

In about two months, Lambda was reported to have signed the $35 billion Anthropic deal, more than tripled its order backlog, and opened talks for a pre-IPO round of up to $4 billion at a $14.5 billion pre-money valuation led by Blackstone and Coatue. It also closed a $1 billion debt financing and announced an Oklahoma data center. The round is still a reported negotiation, not a closed deal. Until a round closes it is only talk, so we look at who is really getting funded, and by whom, in our AI infrastructure report.

Fireworks combined a $1.505 billion round with revenue and token traffic multiplying, then was reported in October 2026 to be in early talks at a valuation of $30 billion to $40 billion, even as it dealt with a recent security incident. Together added $800 million, signed the HUMAIN project, and kept shipping products. Baseten’s token volume grew about 40 times. Runpod passed $325 million in recurring revenue and 1.2 million developers.

Crusoe moved fastest among the physical builders: a $3.9 billion Series F, a reported $13 billion Jane Street contract, and new partners in Perplexity and Thinking Machines Lab. Nscale raised $3.36 billion and filed for an IPO, though its operating capacity still trails the size of its contract book. Fluidstack raised $1.5 billion led by Jane Street at a reported $18 billion valuation in September 2026, and the Pentagon was reported the same month to be in talks to lend it about $5 billion. Anthropic still shapes most of its story.

Which AI cloud startups are actually ahead?

Lambda is narrowly ahead overall today, Fireworks AI is the closest challenger and leads inference, and Together AI is third.

We give the top position to Lambda because it now has the largest reported revenue in the field, a contract book of about $50 billion, and clusters customers use today. The lead is narrow. Its revenue and backlog come from press reports, and two buyers, Anthropic and Microsoft, carry much of the contract value.

Fireworks ranks second, close behind. It has the cleanest company-stated proof: more than $1 billion in annualized revenue, more than 40 trillion daily tokens, and almost all traffic on customized models. Together ranks third. It covers more of the model lifecycle, but bookings give less certainty than recognized revenue, and its HUMAIN forecast depends on a data center not yet built.

Baseten ranks fourth after a strong year of growth and a strong customer list. Crusoe comes fifth, ahead of Nscale, because more of its physical system is visible and operating. Runpod ranks sixth: its more than $325 million in recurring revenue tops Nscale’s first-half revenue on an annualized basis, about $281 million, and it leads in developer reach and capital efficiency. Nscale ranks seventh; its IPO filing confirms both its scale and its delivery risk. Fluidstack ranks eighth. The ranking weighs revenue and customers, not valuation, which is why a reported $18 billion valuation does not lift it: Anthropic has given it an enormous opportunity, with less evidence so far of a diversified business.

The ranking can change under clear conditions. Fireworks retakes first if its next disclosure shows revenue growing faster than Lambda’s, or if Lambda’s reported figures prove lower when it files to go public. Together moves up if bookings and the HUMAIN capacity turn into recognized revenue. Crusoe or Nscale can enter the top three once contracted capacity runs and produces repeatable cloud economics.

Rank	Startup	Why it is here
1	Lambda	Most mature private GPU cloud, a reported $35B Anthropic deal, and a reported $50B backlog; figures are not company-stated
2	Fireworks AI	Leader in inference, with the best company-stated mix of revenue, production traffic, growth, and customized workloads
3	Together AI	Similar commercial scale, broader platform, and strong customers; bookings and forecasts are less conclusive than recurring revenue
4	Baseten	Strong growth and customer traction, with less public absolute data
5	Crusoe	Strongest physical advantage, a $3.9B Series F, and a widening customer base
6	Runpod	More than $325M in ARR, 1.2 million developers, and the best funding efficiency in the field
7	Nscale	Largest future commitments and an IPO filing, with about 55MW operating and heavy losses
8	Fluidstack	Huge Anthropic opportunity and a reported $18B valuation, with customer concentration and execution still defining the case