AI inference: which startup is ahead?

Which AI inference startups actually belong in this comparison?

Eight startups belong in the core AI inference comparison today: Fireworks AI, Baseten, Together AI, Groq, fal, Modal, DeepInfra, and OpenRouter.

We look at companies that help developers run models in production, not every business that sells GPUs or designs AI chips. The wider field has more than 20 credible names, but many sit in a different layer of the market.

Fireworks, Baseten, and Together are the most direct competitors. Each helps customers deploy open or custom models, tune their performance, and handle large production workloads.

Groq qualifies because GroqCloud runs inference directly on its own hardware. Since the Nvidia licensing deal, it describes itself as an inference cloud, and in August 2026 it became an Nvidia Cloud Partner.

Fal focuses on image, video, audio, and 3D generation. Modal sells broader serverless AI compute, with inference as one major workload next to sandboxes and training. DeepInfra competes on low-cost model APIs. OpenRouter sits above the providers and routes requests across more than 400 models.

OpenRouter stays in the table, with a caveat. Stripe announced in August 2026 that it would acquire the company. OpenRouter says its product and routing stay neutral, but it is no longer an independent startup once the deal closes.

We leave out CoreWeave, Nebius, Lambda, and other neoclouds because they sell a much broader mix of training and compute. RunPod and Replicate remain relevant developer platforms, but their public disclosures give us less to compare.

Cerebras is a public company, so it no longer fits a startup ranking. Etched and other chip companies are still too early commercially, with little public evidence of repeated, large production deployments. Gimlet Labs, which raised $300 million in September 2026 to run inference across mixed chips including Cerebras, is one to watch, but it has published no revenue figure we can compare.

Funding figures are messy in this category. Databases count secondary sales, strategic commitments, and tranches differently. We use rounded cumulative amounts and show ranges where records disagree.

Startup	Main position in AI inference	Approximate cumulative funding
Fireworks AI	Custom-model training and production inference	$1.83 billion (new round at $30 billion to $40 billion reported in talks)
Baseten	Managed, hybrid, and self-hosted model deployment	More than $2 billion
Together AI	Full-stack open-model cloud	About $1.2 billion
Groq	Specialized-hardware inference cloud	About $2.75 billion in major disclosed rounds
fal	Generative-media inference	$337 million to about $590 million, depending on the database
Modal	Serverless AI compute, sandboxes, and inference	About $1.2 billion
DeepInfra	Low-cost hosted model APIs	At least $125 million
OpenRouter	Multi-model routing and gateway (being acquired by Stripe)	About $154 million

Is Fireworks AI still ahead now that Modal has hit $1 billion?

Yes. Fireworks AI is still ahead overall, but Modal now matches its headline revenue and Baseten is close enough to challenge it.

Fireworks has the best mix of disclosed revenue, production traffic, custom-model usage, customers, and technical breadth. In its July 2026 Series D announcement, it said annualized revenue had passed $1 billion and daily volume had exceeded 40 trillion tokens.

The two figures matter together. Heavy traffic can come from cheap calls, and annualized revenue can include contracts that are only starting. Fireworks shows both commercial scale and active use, and over 95% of its tokens come from models specialized on customer data. Heavy traffic from cheap calls can flatter a startup, so we separate who has real customers from who only promises in our AI infrastructure report.

Modal changed the picture in October 2026. It said it had crossed $1 billion in annualized run rate, more than triple its level in May. But about $300 million of that comes from sandboxes for AI agents, and Modal does not break out how much is inference.

So Modal has caught up on revenue, not on inference. Its business is broader: sandboxes, GPU clusters for training, and an inference product called Endpoints. Fireworks' $1 billion is almost entirely model serving and post-training.

Baseten is the closest pure-play challenger. It reported twentyfold revenue growth and fortyfold inference-volume growth over the year to June 2026. Its customers include Cursor, Notion, Harvey, HubSpot, OpenEvidence, Abridge, and Decagon.

Together belongs in the leading group, though its business is harder to read. It sells inference, fine-tuning, GPU clusters, storage, and large AI factories. Press coverage of its July round put annual bookings above $1.15 billion, with an undisclosed share outside inference.

The structure holds: Fireworks leads, Baseten and Together chase, and Modal has moved from outsider to a serious rival on scale.

Who has built the biggest AI inference business, and who uses capital best?

Fireworks has the biggest clearly disclosed pure-play inference business. Modal now has one of the best disclosed revenue-to-funding ratios, and a business of similar size.

Fireworks' annualized revenue rose from more than $280 million in late 2025 to above $1 billion by July 2026. Daily token volume climbed from more than 10 trillion to more than 40 trillion over roughly the same period.

Revenue grew by at least 257%, while traffic grew by about 300%. Implied revenue per token fell by roughly 11%, if the company counts traffic the same way. That is a modest decline in a market where inference prices keep falling.

Fireworks has raised about $1.83 billion, so its revenue equals at least 55% of its funding. Business Insider reported in October 2026 that it is in early talks to raise at $30 billion to $40 billion, against $17.5 billion in its July round. That is a proposed price, not a closed round.

Modal has raised about $1.2 billion after its $750 million Series D in October 2026, led by Accel at a $15.75 billion post-money valuation. With more than $1 billion in run rate, its revenue equals roughly 80% of cumulative funding, among the best in this group.

Baseten was reported in June 2026 to have grown its annualized run rate from about $200 million to $600 million. The company confirms its growth rates but not the base. On more than $2 billion raised, that gives an estimated ratio near 30%.

Investors are betting on that growth. Baseten's June round valued it at $13 billion, and Axios reported in September that it had been approached about roughly doubling that to $26 billion. No new round has been announced.

Together reports more than $1.15 billion in annual bookings against about $1.2 billion raised. The ratio looks strong, but bookings can cover future years and include GPU clusters and training.

Groq has raised about $2.75 billion in major rounds, including $650 million in June 2026 and $350 million at a $3.5 billion valuation in August. It publishes no comparable revenue figure.

Fal may be the most efficient of all. The Information reported in September 2026 that it was running at about $800 million a year and had held early talks at a $15 billion valuation or more. Fal has not confirmed either figure.

Fireworks leads on pure inference scale. Modal is among the most capital-efficient and has matched Fireworks' headline revenue, while Baseten carries the biggest bet that future growth will justify its price.

Is Modal now gaining ground faster than Baseten in AI inference?

On recent numbers, yes. Modal is growing fastest overall, while Baseten is still the fastest-growing pure inference challenger.

Modal says usage tripled in the five months between its May and October 2026 rounds and that it has grown more than fifteenfold year over year. Annualized revenue went from more than $300 million to more than $1 billion over the same stretch.

Part of that growth comes from sandboxes, the containers AI agents use to run code. That market is booming, and it is not inference in the narrow sense. Modal's lead in speed of growth is clearest when the race is defined as AI compute rather than model serving.

Baseten's twentyfold revenue growth and fortyfold inference-volume growth, reported in June 2026, still make it the fastest serious pure-play challenger. Its earlier Series E had reported a hundredfold rise in inference volume, so the acceleration has held across consecutive periods.

Baseten is also moving toward Modal's ground. In September 2026 it bought Blaxel, a startup that builds sandboxes for AI agents. The same month it became one of the first open-model providers in OpenAI's enterprise marketplace, with a native integration in Codex.

Traffic at Baseten grew about twice as fast as revenue in the year to June. Revenue per unit of inference seems to have fallen sharply.

Fireworks has grown from a larger base, with revenue and token growth much closer together. Its daily volume rose from about 15 trillion tokens a year earlier to more than 40 trillion.

OpenRouter says it now processes more than 10 trillion tokens a day, against 25 trillion a week in May 2026.

The metrics do not line up exactly, but the picture is readable. Modal is gaining ground fastest, Baseten is closing the gap fastest inside pure inference, and Fireworks is adding the most proven inference business at scale.

Which AI inference startup is strongest at custom models and production scale?

Fireworks is strongest at running custom models in large production environments, with Baseten close behind.

Fireworks serves more than 40 trillion tokens per day, and over 95% come from models adapted to customer data and specific tasks.

Its platform covers serverless inference, dedicated deployments, reinforcement learning, fine-tuning, evaluations, multimodal models, and customer-controlled cloud setups. In August 2026 it made its Training API generally available, pushing further into post-training. Inference platforms stretching into fine-tuning and post-training show which technologies are gaining ground; we look at that in our report on AI infrastructure.

Cursor is a useful example. Fireworks supports production inference and reinforcement-learning workloads behind Cursor's coding models. In August 2026 it also described post-training Kimi K3 with Harvey for long legal tasks.

Baseten offers more deployment flexibility. Customers can use managed cloud, hybrid deployments, self-hosted systems, dedicated endpoints, or embedded engineering teams. Baseten says it runs more than 90 clusters across more than 20 clouds.

That control matters for healthcare, legal, and enterprise customers. OpenEvidence, Abridge, Harvey, Cursor, and Notion all run AI products where downtime or weak data controls would quickly hurt users.

Together covers the broadest model lifecycle, from training and fine-tuning to inference, GPU clusters, and dedicated AI factories. It has commitments for more than 500 megawatts of future capacity, and in October 2026 it added a dedicated B300 inference cluster with IBM Cloud and Nvidia.

Those commitments could become a major edge once the power is installed and used. For now, Fireworks has stronger evidence of current production workloads, while Together has the largest announced pipeline.

Groq runs 13 data centers across four regions. It says capacity is about 54 megawatts today, with plans to pass 200 megawatts in 2027. Its architecture is fast, but it supports a narrower range of unusual custom models than GPU platforms.

Fireworks wins on current custom-model volume. Baseten offers the most deployment control, Together has the broadest stack, and Groq has the widest specialized-hardware footprint.

Who is fastest at AI inference right now?

No single startup is fastest. Groq leads on established real-time models, while Baseten and Fireworks each win on different newer open models.

On Llama 3.3 70B, Artificial Analysis measures Groq at about 347 output tokens per second, with a first token in 0.96 seconds. Together's Turbo endpoint reaches about 68 tokens per second, and DeepInfra about 14.

On GLM-5.3, Baseten's fast endpoint reaches about 282 tokens per second, more than twice Together's 133. Baseten's standard endpoint sits at about 125.

On Kimi K3, Fireworks leads the three main pure-plays at about 112 tokens per second. Baseten follows at about 96, and Together at about 57.

On the latest DeepSeek V4 Pro endpoint, Baseten runs at roughly 115 to 120 tokens per second across two readings, while Together and Fireworks are not currently benchmarked. A smaller provider, Systalyze, leads that test at about 317.

Rankings shift with the model, prompt length, quantization, reasoning setting, traffic, and hardware. They also shift with coverage: when a provider drops off a benchmark, its earlier lead stops being visible.

Groq remains the standout for supported real-time workloads. Among the general platforms, Baseten currently shows the most top results on newer large models, and Fireworks leads on Kimi K3.

Current independent benchmark	Leader among included startups	Closest included rival	Competitive distance
Llama 3.3 70B, output speed	Groq: 347 tokens/second	Together Turbo: 68	Groq is about 5 times faster
GLM-5.3, output speed	Baseten (fast endpoint): 282 tokens/second	Together: 133	Baseten is about 2.1 times faster
Kimi K3, output speed	Fireworks: 112 tokens/second	Baseten: 96	Fireworks is about 17% faster
DeepSeek V4 Pro (latest endpoint), output speed	Baseten: about 115 to 120 tokens/second	DeepInfra: below 70	Together and Fireworks not benchmarked

Which AI inference startup gives customers the best economics?

DeepInfra has the lowest raw API prices, while Together and Fireworks can offer better economics when speed or customization cuts the work required.

Artificial Analysis lists DeepInfra at a blended price of about $0.12 per million tokens for Llama 3.3 70B, the lowest in that comparison. Groq charges about $0.61 for the same model.

On GLM-5.3, DeepInfra lists about $0.45 per million blended tokens. Its cost per benchmark task is about $0.97, against $1.43 for Baseten's standard endpoint and $1.97 for Together.

The trade-off is speed. DeepInfra produces Llama 3.3 output at about 14 tokens per second, where Groq passes 340. That is fine for background jobs, but frustrating in a coding agent or a voice assistant.

Together offers a better balance between speed and price. The company says customers moving to open models often cut inference spending six to twenty times compared with closed-model APIs, and Decagon reportedly cut its costs sixfold. In October 2026 Together launched Together Link, which it says cuts model spend in coding agents by more than 50%.

Fireworks focuses on adapting smaller models to specific jobs. A specialized model can use fewer parameters, produce shorter outputs, and need fewer retries. On Kimi K3, its blended price is about $2.31 per million tokens, close to Together's $2.08 but at nearly twice the speed.

DeepInfra is no longer small. In September 2026 it said its annualized run rate had passed $105 million, about 14 times a year earlier, with 22 trillion tokens served a week.

DeepInfra is the obvious start for cost-sensitive batch work. Together fits customers who want speed and open-model prices together. Fireworks pays off when customization reduces the total amount of inference.

Who has the strongest AI inference customers and contracts?

Fireworks has the strongest overall customer evidence, while Baseten has the densest group of fast-growing AI-native customers and a new channel through OpenAI.

Fireworks previously disclosed more than 10,000 company customers, including Samsung, Uber, DoorDash, Shopify, Notion, Upwork, Cursor, and Harvey.

Several of these involve core production systems. Cursor uses Fireworks for training and inference, and Harvey worked with Fireworks to post-train Kimi K3 for legal work. Innovative Solutions said it moved roughly 90% of its Anthropic inference spending to Fireworks within two weeks.

Baseten's customer list includes Cursor, Notion, Lovable, Harvey, HubSpot, OpenEvidence, Abridge, Decagon, and Parallel. In September 2026 it joined OpenAI's B2B marketplace, so OpenAI enterprise customers can apply existing OpenAI commitments to open models served by Baseten.

That deal matters more than a logo. It puts Baseten inside the buying process of companies that already spend with OpenAI. No financial terms were disclosed.

Together serves thousands of customers, including Cognition, Decagon, ElevenLabs, Cursor, and Suno. It can expand from inference into training and reserved GPU capacity in the same account.

Modal names DoorDash, Cognition, Ramp, Linear, Runway, and Physical Intelligence among its customers. Many use it for agents and training as well as inference.

Groq's strongest relationships are strategic. It provides inference for Bell Canada's sovereign AI network, has worked with Meta on the Llama API, and became an Nvidia Cloud Partner in August 2026. It discloses few contract values.

Fireworks leads on breadth and depth of usage. Baseten follows closely on customer quality and now has the strongest new distribution deal, while Together has the most room to grow spending inside each account.

Does OpenRouter still lead developer distribution after the Stripe deal?

Yes. OpenRouter is still the largest developer distribution layer, but it is joining Stripe and will no longer be an independent startup. Groq and fal follow.

Stripe and OpenRouter announced the deal on 19 August 2026. The price was not disclosed. Bloomberg reported more than $7 billion, the New York Times $7.5 billion, and Axios above $8 billion, so reports disagree on the exact figure. Estimates from $7 billion to over $8 billion show how slippery a company's worth can be; we cover valuations in our AI infrastructure report.

OpenRouter says more than 10 million developers and companies use it, across more than 400 models. It processes more than 10 trillion tokens a day.

Developers use one interface to reach many model makers and providers. OpenRouter handles routing, fallbacks, spending limits, provider choice, and data policies.

Most of that traffic runs on other companies' infrastructure, so OpenRouter is a marketplace and control layer more than a compute provider. It says routing will stay neutral under Stripe. Providers will watch whether that holds.

Groq serves more than six million developers directly, up from five million earlier in 2026, on infrastructure it controls.

Fal's site says more than 1.5 million developers use the platform, with more than 1,000 production-ready media models.

Fireworks disclosed hundreds of thousands of developers in late 2025, far fewer than OpenRouter or Groq. Yet it generates more revenue than either has confirmed, which points to a smaller audience running much heavier production work.

OpenRouter wins distribution, and that position is passing to Stripe. Fireworks wins monetization depth, while Groq combines broad reach with inference it runs itself.

Which AI inference startup has an advantage competitors cannot easily copy?

Fireworks now owns the cleaner moat. Groq still has the hardest technology to rebuild, but Nvidia now sells hardware based on it.

Building Groq's processor, compiler, networking, and deployment know-how took years. Recreating that across 13 data centers takes far more capital and engineering than launching a GPU-based API. Years of chip and compiler work are hard to match; we look at how far the challengers are behind in our AI infrastructure report.

The Nvidia deal weakened that edge. Nvidia licensed Groq's inference technology, and founder Jonathan Ross, president Sunny Madra, and other staff joined Nvidia. In August 2026 Groq said it was among the first to bring Nvidia's Groq 3 LPX systems to market, so the architecture is no longer exclusive to GroqCloud.

Fireworks has a broader, software-heavy advantage. Its founding team helped build PyTorch, and the company combines kernels, model optimization, training, evaluations, serving, and customer-specific operating data.

Each specialized deployment can hold proprietary data, custom evaluations, post-training, routing choices, and performance history. Rivals can copy single features faster than they can rebuild years of embedded customer work.

Together has a strong research moat. FlashAttention became part of the wider AI stack, and its team keeps working on kernels, compilers, and inference. Its 500-plus megawatts of committed capacity could add a physical moat if it fills.

Modal's edge is its programming model and its sandbox scale. It says it ran one million concurrent containers in an internal test. Baseten's purchase of Blaxel shows rivals see that layer as worth owning.

Groq still owns the deepest hardware barrier, but shares it with Nvidia. Fireworks has the strongest current mix of technology, integration, scale, and team continuity.

Which startup leads each part of the AI inference market?

Fireworks leads general production inference, while six rivals hold credible positions in narrower parts of the category.

Fireworks is strongest where custom models, production reliability, and large recurring workloads meet. Its closest rivals are Baseten, which offers more deployment control, and Together, which covers more of the model lifecycle.

No one owns raw speed on large open models. Baseten and Fireworks each win on different models in current Artificial Analysis tests, so customers should benchmark the model they plan to run.

Groq owns the specialized low-latency hardware position. Customers who need fast generation on supported models may pick Groq even when another platform has a broader catalog.

Fal is the clearest generative-media specialist across image, video, audio, and 3D, and a reported $800 million revenue pace suggests that niche is now large.

Modal leads programmable serverless AI compute and agent sandboxes. Developers can mix inference, agents, code execution, training, and batch jobs in one Python-centered environment.

DeepInfra owns the low-price position on many open models. OpenRouter leads model aggregation and routing, and is joining Stripe.

The shape holds: Fireworks leads the central production contest, Baseten and Together stay within reach, Modal has grown into a scale rival from the compute side, and the others get much stronger when the comparison narrows to one workload.

How trustworthy are the AI inference comparisons?

The ranking is more reliable than the exact gaps between the companies.

Private startups choose which numbers to publish, and they rarely use the same definitions. Annualized revenue, bookings, contract value, token volume, developers, and customers can describe very different things.

Fireworks, Modal, and DeepInfra disclose annualized run rates directly. Baseten confirms its growth rates but not its base; the $600 million figure comes from press reports. Together's scale is reported mainly through bookings. Fal's revenue figure comes from a single press report.

Sources sometimes disagree. Baseten announced its June round at $13 billion, while some reports described two tiers priced at $11 billion and $13 billion. Reports of Stripe's price for OpenRouter range from more than $7 billion to above $8 billion.

Usage needs care too. Fireworks counts tokens served on its platform. OpenRouter counts traffic routed across providers. Fal often measures media generations or calls.

Artificial Analysis gives a stronger basis for speed, since it tests providers on the same models and prompts. But results change with capacity, hardware, quantization, and which providers are covered at a given time.

We have more confidence in Fireworks' overall lead because revenue, traffic, customers, custom-model usage, and funding all point the same way.

Which AI inference startups are actually ahead?

Fireworks AI is still ahead overall, followed by Baseten and Together AI in a close top three, with Modal now right behind them.

Fireworks earns first place because its strongest metrics describe a working production business. Annualized revenue has crossed $1 billion, most usage comes from specialized customer models, and relationships such as Cursor and Harvey reach deep into customers' products.

Baseten takes second place as the most dangerous pure-play challenger. Twentyfold revenue growth and fortyfold volume growth, hybrid and self-hosted options, the Blaxel deal, and its place in OpenAI's marketplace widen the market it can reach.

On reported revenue, Fireworks may be about 1.7 times larger. Investors are pricing Baseten as if it will close that gap: the reported $26 billion approach would put it above Fireworks' last closed round, though below the range Fireworks is now discussing.

Together ranks third. It has the broadest open-model stack, more than 500 megawatts of committed capacity, and new enterprise inference capacity with IBM Cloud and Nvidia. It could move higher once it separates inference revenue from its broader infrastructure business.

Modal rises to fourth. More than $1 billion in run rate and fifteenfold yearly growth make it the biggest mover of the year. It stays behind the top three because sandboxes and training make up a large, undisclosed share of that revenue.

Groq falls to fifth. Six million developers, 13 data centers, and fresh capital give it real scale, but the Nvidia deal took its founders and now lets Nvidia sell its architecture.

Fal takes sixth overall and first in generative-media inference. Its reported $800 million revenue pace could justify a higher place, but the figure is unconfirmed and its work is narrower. OpenRouter ranks seventh: it owns valuable distribution, is being acquired by Stripe, and controls little of the underlying compute. DeepInfra ranks eighth, leading on raw price with less evidence of deep custom deployments.

Fireworks is still the company to beat. The ranking could change if Baseten turns growth into comparable revenue, Modal shows that most of its run rate is inference, or Together fills its capacity with inference demand.

Rank	Startup	Why it ranks here
1	Fireworks AI	Best combination of inference revenue above $1 billion, custom-model usage, production scale, customers, and momentum
2	Baseten	Fastest-growing pure-play challenger, with strong AI-native customers, flexible deployment, and a new OpenAI marketplace channel
3	Together AI	Broadest full-stack open-model platform and the largest committed capacity pipeline
4	Modal	More than $1 billion in run rate and the sharpest recent growth, with inference only one part of a broader compute and sandbox business
5	Groq	Strong specialized hardware, broad developer reach, and global infrastructure, with its architecture now shared with Nvidia
6	fal	Clear leader in image, video, audio, and 3D inference, although its category is narrower
7	OpenRouter	Dominant multi-model routing layer, being acquired by Stripe, and relying on outside inference providers
8	DeepInfra	Strongest low-cost provider in several benchmarks, with less visible enterprise depth and customization