AI inference: which startup is ahead?
In our AI infrastructure market deck, you will find everything you need to understand the market
SUMMARY
Fireworks AI is ahead overall today, with Baseten and Together AI close enough to make the lead contestable rather than comfortable.
Fireworks has the strongest mix of disclosed revenue, daily token volume, custom-model usage, production customers, and technical breadth. No rival currently matches that combination with equally clear operating evidence.
Baseten is the fastest-moving challenger. Its reported twentyfold revenue growth and fortyfold inference-volume growth suggest it is closing the gap quickly, although the undisclosed revenue base makes the remaining distance hard to measure.
Together AI looks strongest when the comparison shifts from current pure-play inference revenue to open-model performance, full-stack infrastructure, and future capacity. Its reported bookings are large, but they also include products beyond inference.
The market does not have one universal winner. Groq leads specialized low-latency hardware, fal leads generative-media inference, Modal leads programmable serverless AI compute, DeepInfra leads on raw price, and OpenRouter leads the routing and distribution layer.
Fireworks’ most important advantage is not simply traffic. More than 95% of its reported tokens come from models adapted to customer data and specific jobs, which points to deeper integrations and higher switching costs than a generic model API.
Developer reach and commercial depth are separating. OpenRouter and Groq have much larger disclosed developer audiences, while Fireworks appears to monetize a smaller base through heavier production workloads.
Performance leadership changes by model. Together currently leads several newer large open-model benchmarks, while Groq remains unusually fast on supported real-time workloads such as Llama 3.3 70B.
Capital efficiency is harder to judge than the funding tables suggest. Modal has the strongest simple revenue-to-funding ratio, but bookings, annualized revenue, contract value, routed traffic, and infrastructure commitments are not equivalent measures.
The ranking is therefore more reliable than the exact gaps. Fireworks is the company to beat, Baseten is gaining fastest, Together has the broadest expansion case, and every other startup becomes much stronger when the market is narrowed to a specific workload.
This market map, featured in our AI infrastructure market deck, highlights top companies and startups in the AI infrastructure market
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 are looking at companies that help developers run models in production, rather than every business selling GPUs or designing AI chips. The wider universe contains 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, optimize their performance, and handle large production workloads.
Groq also qualifies because GroqCloud directly runs inference on its specialized hardware. Its business now resembles an inference cloud more than a conventional chip supplier.
Fal concentrates on image, video, audio, and 3D generation. Modal provides broader serverless AI compute, with inference as one major workload. DeepInfra competes through low-cost model APIs. OpenRouter sits above the providers and routes requests across more than 400 models.
We exclude CoreWeave, Nebius, Lambda, and other neoclouds from the main ranking because they sell a much broader mix of training and compute infrastructure. RunPod and Replicate remain relevant developer platforms, although their latest public disclosures give us less material for a detailed comparison.
Cerebras has moved into the public markets, so it no longer fits a startup ranking. Etched and several other chip companies remain too early commercially. Etched has announced a working chip and more than $1 billion in customer demand, but public evidence of repeated, large-scale production deployments remains limited.
Funding figures are especially messy in this category. Databases sometimes count secondary transactions, strategic commitments, or separate tranches differently. We use rounded cumulative amounts and show ranges where the available records disagree.
| Startup | Main position in AI inference | Approximate cumulative funding |
|---|---|---|
| Fireworks AI | Custom-model training and production inference | $1.83 billion |
| 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 $2.4 billion in major disclosed rounds |
| fal | Generative-media inference | $337 million to $400 million |
| Modal | Serverless AI compute and inference | About $466 million |
| DeepInfra | Low-cost hosted model APIs | At least $125 million |
| OpenRouter | Multi-model routing and gateway infrastructure | About $154 million |
Is any AI inference startup clearly ahead today?
Fireworks AI is ahead overall today, although Baseten and Together AI remain close enough to challenge it.
Fireworks currently has the best combination of disclosed revenue, production traffic, custom-model usage, customers, and technical breadth. Its latest financing announcement says annualized revenue has passed $1 billion and daily volume has exceeded 40 trillion tokens.
Those figures are more convincing together than either would be alone. High traffic can come from low-value calls, while high annualized revenue can include contracts that are only starting. Fireworks shows both commercial scale and active usage.
More than 95% of its tokens reportedly come from models adapted to customer data and specific jobs. That suggests its largest customers have moved well beyond testing interchangeable public APIs.
Baseten is the closest pure-play challenger. Its latest announcement reports twentyfold revenue growth and fortyfold inference-volume growth over the preceding year. It also serves Cursor, Notion, Harvey, HubSpot, OpenEvidence, Abridge, and Decagon.
Together belongs in the same leading group, though its business is harder to separate. It sells inference, fine-tuning, GPU clusters, storage, and large AI factories. Press coverage around its latest round put annual bookings above $1.15 billion, but an undisclosed share comes from products outside inference.
The current structure is fairly clear: Fireworks leads, Baseten and Together form the chasing group, and the remaining startups lead narrower parts of the market.
If you want more recent data on this point, please see our latest AI infrastructure market report.
As this chart shows, and as featured in our AI infrastructure market deck, search interest in AI infrastructure has risen sharply
Who has built the biggest AI inference business, and who uses capital best?
Fireworks has built the biggest clearly disclosed pure-play inference business, while Modal currently shows the strongest simple revenue-to-funding ratio.
Fireworks’ annualized revenue rose from more than $280 million in late 2025 to above $1 billion. 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%. The implied revenue per token fell by roughly 11%, assuming the company counted traffic consistently. That is a relatively modest decline in a market where inference prices keep falling.
Baseten may now generate around $600 million in annualized revenue, according to outside estimates. At that level, Fireworks would be about 1.7 times larger. Baseten has raised more than $2 billion, giving it an estimated revenue-to-funding ratio near 30%.
Together reportedly has more than $1.15 billion in annual bookings against approximately $1.2 billion raised. The comparison looks strong, but bookings can cover future years and include GPU clusters, training, and other infrastructure.
Modal has raised about $466 million and reports more than $300 million in annualized revenue. Its revenue equals roughly 64% of cumulative funding, compared with at least 55% for Fireworks.
Groq has raised around $2.4 billion across its major disclosed rounds but does not publish a comparable revenue figure. Its developer base and infrastructure are large, yet we cannot judge how efficiently they convert into recurring business.
Fal may also be highly efficient if external estimates placing its annualized revenue near $400 million are accurate, but the company has not confirmed that number directly.
Fireworks leads on commercial scale. Modal leads the crude capital-efficiency comparison, while Baseten is making the largest bet that exceptional future growth will justify its funding.
Which AI inference startup is gaining ground fastest right now?
Baseten is gaining ground fastest in percentage terms, while Fireworks continues to add the most proven business at scale.
Baseten’s twentyfold revenue growth and fortyfold inference-volume growth make it the fastest serious challenger. Its earlier Series E announcement had already reported a hundredfold increase in inference volume during the prior year, so the acceleration has continued across consecutive periods.
Traffic grew about twice as quickly as revenue in the latest period. Average revenue per unit of inference therefore appears to have fallen sharply, although large-customer discounts, cheaper models, and lighter requests may explain part of the decline.
Fireworks has expanded from a larger base, with recent revenue and token growth remaining much closer together. It appears to be preserving more of its implied revenue per token while scaling.
OpenRouter increased weekly volume from 5 trillion to 25 trillion tokens in six months. DeepInfra says token volume has grown twenty-fivefold since its Series A, while Modal’s annualized revenue grew fivefold between September 2025 and May 2026.
Groq expanded from more than two million developers to more than five million in less than a year.
The metrics are not directly interchangeable, but the competitive picture is still readable. Baseten is closing the gap fastest, Fireworks is adding the most visible commercial scale, and Modal and OpenRouter are the strongest outsiders by recent acceleration.
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, shows annual VC investment in AI infrastructure startups
Which AI inference startup is strongest at custom models and production scale?
Fireworks is currently 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% reportedly come from models adapted to customer data and specific tasks. That works out to at least 38 trillion specialized tokens daily.
Its platform covers serverless inference, dedicated deployments, reinforcement learning, fine-tuning, evaluations, multimodal models, and customer-controlled cloud configurations. It also operates across more than 18 regions and eight cloud providers.
Cursor is a useful example. Fireworks supports production inference and reinforcement-learning workloads behind Cursor’s coding models, placing it inside both model improvement and serving.
Baseten offers greater deployment flexibility. Customers can use managed cloud infrastructure, hybrid deployments, self-hosted systems, dedicated endpoints, or embedded engineering teams.
That control is valuable for healthcare, legal, and enterprise customers. OpenEvidence, Abridge, Harvey, Cursor, and Notion all rely on AI systems where downtime, poor model migrations, or weak data controls would quickly affect users.
Together covers an even broader model lifecycle, including training, fine-tuning, evaluations, storage, inference, GPU clusters, and dedicated AI factories. It has also secured commitments for more than 500 megawatts of future capacity.
Those commitments could become a major advantage once the power is installed and utilized. For now, Fireworks has stronger evidence of current production workloads, while Together has the largest announced expansion pipeline.
Groq operates 13 data centers across four major regions and plans to approach 200 megawatts by the end of 2027. Its specialized architecture delivers strong performance, though it supports a narrower range of unusual custom models than flexible GPU platforms.
Fireworks wins through current custom-model volume. Baseten offers the most deployment control, Together has the broadest stack, and Groq owns the widest specialized-hardware footprint.
Who is fastest at AI inference right now?
Together AI currently wins the newest large open-model tests, while Groq remains exceptionally fast on several established low-latency workloads.
On DeepSeek V4 Pro at high reasoning effort, Artificial Analysis measures Together at about 339 output tokens per second. Fireworks reaches roughly 183, while Baseten reaches around 140. Together is about 85% faster than Fireworks and more than 2.4 times faster than Baseten on that endpoint.
Together also delivers the first answer token in 6.84 seconds. Fireworks takes 12.73 seconds, and Baseten takes around 15 seconds.
The same pattern appears on GLM-5.2 at maximum reasoning effort. Together reaches approximately 449 tokens per second, compared with 346 for Baseten and 339 for Fireworks.
Groq shines on models that fit its architecture particularly well. Artificial Analysis currently measures Groq at about 302 tokens per second on Llama 3.3 70B, with a time to first token of 0.95 seconds.
Rankings change with the model, prompt length, quantization, reasoning configuration, traffic load, and hardware generation. Cerebras, now a public company, shows how high the ceiling can go by serving gpt-oss-120B at roughly 1,900 tokens per second.
Together currently has the strongest performance across newer large open models. Groq remains the standout option for supported real-time workloads.
| Current independent benchmark | Leader among included startups | Closest included rival | Competitive distance |
|---|---|---|---|
| DeepSeek V4 Pro, high reasoning, output speed | Together: 338.5 tokens/second | Fireworks: 182.6 | Together is about 85% faster |
| DeepSeek V4 Pro, first-answer latency | Together: 6.84 seconds | Fireworks: 12.73 | Together is about 46% quicker |
| GLM-5.2, maximum reasoning, output speed | Together: 449.3 tokens/second | Baseten: 346.0 | Together is about 30% faster |
| Llama 3.3 70B, output speed | Groq: 301.9 tokens/second | No included startup above 100 | Groq has a wide lead on this model |
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, shows why CoreWeave is winning in AI infrastructure
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 reduces the work required.
Artificial Analysis currently places DeepInfra among the cheapest hosts for several large open models. Its blended price for GLM-5.2 is approximately $0.61 per million tokens, the lowest among 16 tracked providers.
DeepInfra also charges around $0.12 per million blended tokens for Llama 3.3 70B, roughly 8.6 times below the most expensive provider in the comparison.
The trade-off is speed. DeepInfra generates Llama 3.3 output at about 13 tokens per second in the same benchmark where Groq exceeds 300. That may be fine for background processing but frustrating in a coding agent or voice assistant.
Together provides a stronger balance between speed and price. The company says customers using open models often cut inference spending by six to twenty times compared with closed-model APIs. Decagon reportedly reduced its inference costs sixfold after moving workloads to Together.
Fireworks focuses on adapting smaller models to specific jobs. A specialized model can use fewer parameters, produce shorter outputs, and require fewer retries.
DeepInfra is the obvious starting point for cost-sensitive batch workloads. Together looks strongest when customers need both speed and open-model pricing. Fireworks becomes more attractive when customization can reduce the total amount of inference required.
Who has the strongest AI inference customers and contracts?
Fireworks has the strongest overall customer evidence today, while Baseten has the densest group of fast-growing AI-native customers.
Fireworks previously disclosed more than 10,000 company customers, including Samsung, Uber, DoorDash, Shopify, Notion, Upwork, Cursor, and Harvey.
Some of these relationships clearly involve important production systems. Cursor uses Fireworks for model training and inference. Innovative Solutions said it shifted roughly 90% of its Anthropic inference spending to Fireworks within two weeks, then scaled to between 4 billion and 10 billion tokens per month.
Baseten’s customer list includes Cursor, Notion, Lovable, Harvey, HubSpot, OpenEvidence, Abridge, Decagon, and Parallel. Many of these companies sell products where model quality and uptime directly shape the user experience.
Abridge and OpenEvidence add credibility because healthcare AI faces tighter privacy, reliability, and procurement requirements. Harvey provides a similar test in legal work.
Together serves thousands of customers, including Cognition, Decagon, ElevenLabs, Cursor, and Suno. It can also expand from inference into training and reserved GPU capacity within the same account.
Groq’s strongest relationships are more strategic. It provides inference for Bell Canada’s sovereign AI network, has worked with Meta on the official Llama API, and has expanded through infrastructure partnerships in the Middle East and Asia.
Groq has disclosed few contract values or minimum commitments, so those relationships remain harder to compare with active production workloads.
Fireworks leads through breadth and depth of usage. Baseten follows closely on customer quality, while Together has the broadest opportunity to increase spending inside each account.
This chart, included in our AI infrastructure market deck, shows annual funding in AI infrastructure startups
Which AI inference startup has the strongest developer distribution?
OpenRouter owns the largest developer distribution layer today, followed by Groq and fal.
OpenRouter says more than eight million developers use its platform across over 400 models. Weekly volume rose fivefold in six months, reaching 25 trillion tokens.
Developers use one interface to access many model companies and infrastructure providers. OpenRouter handles routing, fallbacks, spending limits, provider selection, and data policies.
A large share of that traffic ultimately runs on infrastructure operated by other companies. OpenRouter therefore resembles a marketplace and control layer more than a direct compute provider.
Groq directly serves more than five million developers. That base grew from two million in less than a year, and the requests run on Groq-controlled infrastructure.
Fal’s current site says more than 1.5 million developers use the platform. The company also offers more than 1,000 production-ready media models.
Fireworks disclosed hundreds of thousands of developers in late 2025, far below OpenRouter or Groq. Yet it now generates more revenue than either company has publicly confirmed, suggesting a smaller audience with much larger production workloads.
OpenRouter wins the distribution race. Fireworks wins monetization depth, while Groq combines broad reach with directly executed inference.
Which AI inference startup has an advantage competitors cannot easily copy?
Groq has the hardest technology to reproduce, while Fireworks currently owns the cleaner business moat.
Building Groq’s processor, compiler, networking system, and deployment expertise took years. Recreating that architecture across 13 data centers requires far more capital and engineering work than launching a GPU-based API.
The Nvidia licensing agreement reduced the exclusivity of that advantage. Nvidia gained access to Groq’s inference technology, while founder Jonathan Ross, president Sunny Madra, and other team members joined Nvidia. GroqCloud continues independently under new leadership.
Fireworks has a broader software-heavy advantage. Its founding team helped create PyTorch, and the company combines kernels, model optimization, training, evaluations, serving, and customer-specific operating data.
Each specialized deployment can include proprietary data, custom evaluations, post-training, routing decisions, and performance history. Competitors can copy individual features more quickly than they can reproduce years of embedded customer work.
Together also has a strong research moat. FlashAttention became part of the wider AI infrastructure ecosystem, and its team continues to work on kernels, compilers, inference, and model optimization.
OpenRouter is developing a data advantage through its view of model selection, provider performance, prices, outages, and workload patterns.
Groq owns the deepest hardware barrier. Fireworks has the strongest current combination of technology, integration, scale, and organizational continuity.
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, compares the main business model options for AI cloud infrastructure providers
Which startup leads each part of the AI inference market?
Fireworks leads general production inference today, while six rivals already own credible positions in narrower parts of the category.
Fireworks is strongest where custom models, production reliability, and large recurring workloads come together. Its closest competitors are Baseten, which offers more deployment control, and Together, which covers more of the model lifecycle.
Together currently leads high-performance open-model serving. Its recent benchmark results on DeepSeek V4 Pro and GLM-5.2 stand out, and its research team keeps turning systems work into production infrastructure.
Groq owns the specialized low-latency hardware position. Customers that value rapid generation on supported models may prefer Groq even when another platform offers a broader catalog.
Fal is the clearest generative-media specialist across image, video, audio, and 3D generation.
Modal leads programmable serverless AI compute. Developers can mix inference, agents, code execution, training, reinforcement learning, and batch jobs inside the same Python-centered environment.
DeepInfra owns the low-price position on many open models. OpenRouter leads model aggregation and routing.
The market has developed a recognizable shape: Fireworks leads the central production contest, Baseten and Together remain within striking distance, and the other companies become 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 identical definitions. Annualized revenue, bookings, contract value, token volume, inference calls, developers, and customers can describe very different realities.
Fireworks and Modal directly disclose annualized revenue. Baseten confirms its growth rate but leaves the current base undisclosed. Together’s commercial scale is reported mainly through bookings. Estimates for fal and Baseten come from external company-data services.
Usage also needs careful interpretation. Fireworks counts tokens served through its platform. OpenRouter counts traffic routed across providers. Groq reports trillions of tokens per week without giving an exact current total. Fal often measures media generations or inference calls.
Artificial Analysis gives us a stronger basis for performance comparisons because it tests providers on the same models and prompt configurations. Even those results can change when providers alter capacity, hardware, quantization, or routing.
We have more confidence in Fireworks’ overall lead because revenue, traffic, customers, custom-model usage, funding, and benchmark performance all point in the same direction.
The distance between Baseten and Together remains harder to judge. Baseten provides stronger growth and customer evidence, while Together provides better current benchmarks, reported bookings, and a much larger capacity pipeline.
This chart, featured in our AI infrastructure market deck, shows the share of revenue generated by each customer segment in the AI infrastructure market
Which AI inference startups are actually ahead?
Fireworks AI is ahead overall today, followed by Baseten and Together AI in a close top three.
Fireworks earns first place because its strongest metrics describe a functioning production business. As seen above, annualized revenue has crossed $1 billion, most usage comes from specialized customer models, and relationships such as Cursor reach deep into customers’ products.
Baseten takes second place because it is the most dangerous challenger. Twentyfold revenue growth and fortyfold volume growth show exceptional acceleration, while its hybrid and self-hosted options widen the enterprise market it can address.
Based on external revenue estimates, Fireworks may currently be about 1.7 times larger. Baseten could erase that difference quickly if growth continues without severe margin pressure.
Together ranks third. It currently has the strongest benchmark performance on several major open models and has secured more than 500 megawatts of future capacity. Its platform also covers training, fine-tuning, inference, storage, and GPU clusters.
Together could move higher once it separates inference revenue from its broader infrastructure business. We still cannot tell how much of its reported bookings comes from serving models and how much comes from large compute contracts.
Groq ranks fourth. Five million developers, 13 data centers, differentiated hardware, and sovereign infrastructure relationships give it genuine scale. The Nvidia licensing agreement weakened its independence and created uncertainty around how much technical leadership remains inside the company.
Fal takes fifth place overall and first place in generative-media inference. Modal follows in sixth, with one of the strongest revenue-to-funding profiles. OpenRouter ranks seventh because it owns valuable distribution without controlling most of the underlying compute. DeepInfra ranks eighth, leading on raw price while showing less evidence of deep custom enterprise deployments.
Fireworks has become the company to beat. The ranking could change if Baseten converts growth into comparable revenue, Together fills its capacity pipeline with inference demand, or Groq turns its post-Nvidia structure into a faster-growing cloud business.
| Rank | Startup | Why it ranks here |
|---|---|---|
| 1 | Fireworks AI | Best combination of revenue, custom-model usage, production scale, customers, performance, and current momentum |
| 2 | Baseten | Fastest-growing direct challenger, with strong AI-native customers and flexible enterprise deployment |
| 3 | Together AI | Best current open-model speed, broadest full-stack platform, and the largest future capacity pipeline |
| 4 | Groq | Strong specialized hardware, broad developer reach, and global infrastructure, with greater strategic uncertainty after the Nvidia agreement |
| 5 | fal | Clear leader in image, video, audio, and 3D inference, although its category is narrower |
| 6 | Modal | Large and rapidly growing serverless AI platform, with inference forming one part of a broader compute business |
| 7 | OpenRouter | Dominant multi-model distribution and routing layer, while relying heavily on outside inference providers |
| 8 | DeepInfra | Strongest low-cost provider in several benchmarks, with less visible enterprise depth and customization |
If you want more recent data on this point, please see our latest AI infrastructure market report.
OUR METHODOLOGY
This analysis asks which AI inference startup is ahead by separating overall market leadership from specialist leadership. We assess commercial scale, growth, production adoption, custom-model capabilities, performance, economics, customer quality, developer reach, competitive advantages, and overall market position before bringing the findings together into a final ranking.
We prioritized the most recent operating evidence available in the material reviewed: company disclosures, independent provider benchmarks, customer examples, infrastructure announcements, funding data, pricing pages, and product documentation. No single metric decides the ranking.
Revenue, bookings, token volume, inference calls, developers, customers, and contract value are treated as different measures. We compare them where useful, but we do not assume they are interchangeable.
We also separate pure-play inference evidence from broader infrastructure activity. Together AI, Modal, and Groq all operate businesses that extend beyond a narrow managed-inference product, so their reported scale cannot always be attributed to inference alone.
Performance comparisons rely primarily on Artificial Analysis because it tests providers on the same models and prompt configurations. Those benchmarks are treated as current snapshots rather than permanent rankings, since speed and latency can change with model versions, quantization, routing, traffic, and hardware.
Funding totals are rounded because databases may count secondary transactions, strategic commitments, and separate financing tranches differently. External revenue estimates are identified as estimates and receive less weight than direct company disclosures.
This is not a weighted scoring model. It is a structured editorial assessment that checks whether commercial, technical, customer, and usage evidence point in the same direction, then uses that combined picture to form the final ranking.
Key sources include Fireworks AI’s Series D announcement, Baseten’s Series F announcement, Baseten’s Series E announcement, Together AI’s company blog, Together AI’s documentation, Groq’s growth-funding announcement, fal’s documentation, Modal’s documentation, DeepInfra’s pricing page, OpenRouter’s Series B announcement, OpenRouter’s documentation, Artificial Analysis, Cursor’s engineering blog, Harvey’s newsroom, Abridge’s newsroom, and OpenEvidence’s newsroom.
This chart, included in our AI infrastructure market deck, shows how GPU cloud infrastructure technology has evolved over time
Related blog posts
- GPU efficiency: which startup is ahead?
- AI cloud: which startup is ahead?
- AI chips: which startup is ahead?
