AI Infrastructure Companies: The Map, and Where the Story Is

Most lists of AI infrastructure companies near the top of a Google search are rankings: fastest-growing, most valuable, most promising. AI infrastructure companies build what AI runs on: the chips, the cloud capacity, and the software that serves, monitors and routes models, from Nvidia and CoreWeave to Fireworks AI and Modal. What a ranking doesn't show is where a challenger still has room to be the subject of a story rather than a comparison inside one.

Because that's the actual condition of this market. One company sets the frame, and everyone else gets positioned against it whether they participate or not. A reporter writing about your inference chip is writing, at minimum, a paragraph about why it isn't a GPU.

The seven layers below are a map of where that gravity is strongest and where it thins out.

Key takeaways

• In 2026, two inference platforms said they had passed $1 billion in annualized revenue: Fireworks AI, valued at $17.5 billion in July, and Modal, valued at $15.75 billion in October.
• In December 2025 Nvidia licensed Groq's inference technology and hired its founder and senior leaders, later disclosing $17 billion in consideration. Groq carries on as a cloud business, valued at $3.5 billion in August.
• Compute turned into a financing story: CoreWeave reported third-quarter revenue of $3.66 billion and a $132.5 billion backlog, and in October its unsecured debt was yielding about 15%.

1. Silicon: can AI chip startups compete with Nvidia?

The silicon layer is Nvidia plus a field of AI chip startups that investors kept funding all year. Cerebras went public in May at $185 a share, raising about $6.4 billion including the underwriters' option, and closed its first day of trading up 68%. SambaNova raised $350 million in February and then $1 billion at an $11 billion valuation in July. Positron raised $230 million in February and $875 million at a $5 billion valuation in September. Tenstorrent, a Salient client, closed a $1.5 billion Series E at a $6.2 billion valuation in October. In London, Fractile raised a $220 million Series B led by Accel, Factorial Funds and Founders Fund.

The defining event here was Groq. On Christmas Eve 2025, Groq announced a non-exclusive licensing agreement with Nvidia for its inference technology, and founder and CEO Jonathan Ross, president Sunny Madra and other senior leaders left to join Nvidia. Early reports put the price at about $20 billion; Nvidia later disclosed $17 billion in consideration. Groq said it would carry on as an independent company and GroqCloud would keep running. Bernstein's Stacy Rasgon wrote that structuring it as a license "may keep the fiction of competition alive."

What happened next makes the point more sharply. In August, Groq raised $350 million at a $3.5 billion valuation to build out a cloud business. That's about half its valuation a year earlier, though the company left after the Nvidia deal is a different business, and Groq disputes calling it a down round.

That is the narrative environment. The most visible independent challenger in inference silicon had its leadership absorbed and its technology licensed, without a conventional acquisition headline, and the company left behind now runs on the incumbent's hardware.

Where the story is: not in benchmark superiority, which reads as a claim in a category where everyone claims it. It's in a specific workload where a different architecture genuinely wins, explained technically, carried by a person the audience already trusts. Silicon is the layer where an individual's credibility does more work than the company's messaging, which is why the companies making progress here are the ones with a technologist the market already knows.

Chip launch materials usually lead with performance stats. When I review them, I ask what those numbers actually enable. Does it mean faster drug discovery, or curing disease sooner? Better stats are one thing, but what they unlock is the story people care about, even if that part isn't really press release material.

2. Compute and cloud

CoreWeave is the clearest case of what the compute and cloud layer became. Its third-quarter revenue was $3.66 billion, up 168% from a year earlier, with more than 4.6 gigawatts of contracted power and a $132.5 billion backlog. Crusoe, Lambda, Nebius and the hyperscalers occupy the rest, and the private ones are raising at that scale too: Crusoe announced the first close of a $3.9 billion round at a $30.9 billion valuation in September.

Then, in October, the story became how all of it gets paid for. CoreWeave's unsecured debt was yielding about 15%, and its stock fell about 29% from the end of September through October 21.

Where the story is: power, land, supply and financing. This layer is now an infrastructure, energy and financing story as well as a technology story, which means it's also covered by a different set of reporters than the rest of this list: business, energy, credit markets, and increasingly local press where the data centers get built. The differentiation that lands is access to capacity and the economics of getting it, not the software layer on top.

3. Inference and serving

Inference and serving, the platforms that run models for applications, drew four rounds of $750 million or more in 2026. Together AI closed an $800 million Series C on July 1 at an $8.3 billion valuation and says its annual bookings passed $1.15 billion. Fireworks AI announced a $1.505 billion Series D in mid-July at a $17.5 billion valuation, saying it had passed $1 billion in annualized revenue and serves more than 40 trillion tokens a day. Baseten raised $300 million at $5 billion in January, then $1.5 billion in June at up to $13 billion, and Bloomberg reported at the end of October that it is set to raise about $2 billion more at up to $30 billion. Modal raised $355 million at $4.65 billion in May and $750 million at $15.75 billion in October, by which point it said it had crossed $1 billion in annualized revenue.

Where the story is: this is the hardest layer to differentiate in words, because every company here says the same three things: faster, cheaper, open. Those claims cancel out. As with chips, "millions of tokens per second" is a stat; the story is what a customer can do at that speed that wasn't practical before. What doesn't cancel is customer economics at scale: what a specific workload costs to run in production, what it cost before, and who will say so on the record. The companies in this layer that break out do it on named customers and unit economics, not on latency charts.

4. Agent infrastructure: what is it, and who builds it?

Agent infrastructure is a young layer whose boundaries are still being drawn. Browserbase, E2B, Modal and Northflank sit here, providing the browsers, runtimes and isolated environments agents need to actually do things. The money says the category is real: sandboxes now account for $300 million of Modal's annual revenue run rate, according to Accel, which led its latest round.

Where the story is: category definition, which is the most valuable narrative position available and the one with the shortest window. When we launched Browserbase we didn't pitch a headless browser product. We framed it as AI agents needing a bridge to the legacy internet — a way to explain why the infrastructure had to exist at all — and placed it as a VentureBeat exclusive alongside a $6.5 million seed from Kleiner Perkins. Within eighteen months the company raised a $40 million Series B at a $300 million valuation, partnered with Google on Gemini 2.5 Computer Use, and was serving over a thousand companies including Perplexity and Vercel.

5. Data and vector

The data and vector layer includes Pinecone, Weaviate, Qdrant and pgvector, which is the interesting one, because it's a Postgres extension rather than a company.

Where the story is: honestly, this is the hardest layer right now. General-purpose database platforms are moving in (Databricks bought Neon and Snowflake bought Crunchy Data in 2025, both Postgres companies), and "we're a vector database" has lost the novelty it had in 2023. The companies here that still earn coverage do it by moving up the stack into retrieval quality and agent memory rather than defending the storage layer.

6. Observability, tracing and evals

The observability layer, the tools that trace and evaluate what models do, is a growing field, but its best-known names, Langfuse and Arize (the company behind Phoenix), now both belong to bigger companies. ClickHouse bought Langfuse in January. Dynatrace agreed in August to buy Arize for $915 million and closed the deal in October. Buyers have noticed what this layer is worth, and the acquisitions give it a business-news angle. A technical pitch can also show how these tools catch failures or measure improvement in production.

Where the story is: standards and measurement. This is the layer with the clearest path to owning a recurring story, because the industry has no agreed answer to "how do you know the model is working," and whoever produces credible measurement becomes a source rather than a subject. Original data is the whole play here.

7. Orchestration and deployment

The orchestration and deployment layer covers model routing, serving frameworks like vLLM and SGLang, and the deployment platforms around them. Both frameworks now have companies behind them: vLLM's creators launched Inferact with a $150 million seed round at an $800 million valuation, and RadixArk, founded by SGLang contributors, launched publicly in May with a $100 million seed round at a $400 million post-money valuation. Model routers can send each request to the cheapest model expected to handle it well, which makes this layer quietly strategic, and it still gets covered mostly as funding news rather than on its own terms.

Where the story is: the cost argument, told from the buyer's side. It's the layer best positioned to publish real numbers about what running AI in production actually costs, which is a story the business press wants and rarely gets from a credible operator.

What does this map mean for an AI infrastructure company?

Your layer determines whether your job is category creation or displacement, and those are different media strategies with different targets and different timelines. Agent infrastructure companies are still naming their category and should spend accordingly. Silicon challengers are displacing an incumbent narrative and need technical credibility and a spokesperson more than they need volume.

And the gravity is real. Any company in this market that builds a communications plan pretending the incumbent doesn't define the frame will produce materials that reporters silently rewrite into a comparison anyway. Better to choose the comparison you want.

That includes whose numbers you use. In my view a pitch shouldn't lean on a competitor's figures, because quoting a rival's research makes the rival the authority on the problem you want to own.

Common questions

What are the layers of AI infrastructure? Silicon, compute and cloud, inference and serving, agent runtimes and sandboxes, data and vector storage, observability and evaluation, and orchestration and deployment. Companies frequently span more than one: Modal, for example, sells both inference and agent sandboxes.

Which companies build AI infrastructure? By layer: Nvidia, Cerebras, SambaNova and Tenstorrent in silicon; CoreWeave, Crusoe, Lambda and Nebius in compute; Fireworks AI, Together AI, Baseten and Modal in inference; Browserbase and E2B in agent infrastructure; Pinecone, Weaviate and Qdrant in vector data; Langfuse and Arize, both now acquired, in observability; and Inferact and RadixArk, the companies behind vLLM and SGLang, in orchestration.

Which AI infrastructure companies are growing fastest? No published ranking settles it, because valuations, revenue run rates, bookings and quarterly revenue measure different things. In inference, Fireworks AI was valued at $17.5 billion in July with more than $1 billion in annualized revenue, Modal at $15.75 billion in October after crossing the same mark, Baseten at up to $13 billion in June, and Together AI at $8.3 billion with bookings past $1.15 billion. In compute, CoreWeave's third-quarter revenue rose 168% to $3.66 billion.

Can an AI chip startup compete with Nvidia? On specific workloads, yes, and investors keep backing the attempt: Cerebras went public, and SambaNova, Positron and Tenstorrent each raised at multibillion-dollar valuations in 2026. Groq shows one way it can end: Nvidia licensed its technology and hired its leadership, disclosing $17 billion in consideration, and Groq now runs a cloud on Nvidia systems.

What is agent infrastructure? The runtimes, browsers and isolated environments AI agents use to take actions rather than just generate text. Providers increasingly combine browsers, runtimes and sandboxes, so the boundaries are still being drawn.

Why is it hard to get coverage as an AI infrastructure company? In my experience, because most pitches make claims that every competitor also makes (faster, cheaper, more open) in a market where one incumbent already frames the conversation. Coverage follows specificity: a named customer, a real workload, and a number that survives a follow-up question.

Sources

• Cerebras Systems, Cerebras Systems Announces Closing of Initial Public Offering, GlobeNewswire, May 15, 2026
• Jordan Novet, Cerebras pops 68% in Nasdaq debut, CNBC, May 14, 2026
• AI chip maker SambaNova raises $1B at $11B valuation, 5 months after last mega round, TechCrunch, July 8, 2026
• Positron, Positron AI Raises $230 Million Series B at Over $1 Billion Valuation, Business Wire, February 4, 2026
• Positron AI raises $875m at a $5bn valuation, DatacenterDynamics, September 2026
• Tenstorrent, Tenstorrent Announces a $1.5B Series E at a $6.2B Valuation, October 19, 2026
• Anwesha Pattanaik, Fractile raises $220m in Series B funding round, Tech Monitor, May 14, 2026
• David Faber, Nvidia buying AI chip startup Groq's assets for about $20 billion in its largest deal on record, CNBC, December 24, 2025
• Nvidia-Groq deal is structured to keep 'fiction of competition alive,' analyst says, CNBC, December 26, 2025
• Nvidia, Annual report for fiscal 2026, Groq note, SEC filing, fiscal year ended January 25, 2026
• Rebecca Bellan, Groq raises $350M to fuel its pivot from AI chips to neocloud, TechCrunch, August 17, 2026
• CoreWeave, CoreWeave Reports Strong Third Quarter 2026 Results, November 10, 2026
• Seema Mody, CoreWeave's market confidence erodes as bond yields spike and stock sinks, CNBC, October 21, 2026
• Crusoe, Crusoe Raises $3.9B Series F at a $30.9B Valuation, September 17, 2026
• Julie Bort, Neocloud Together AI raises $800M, leaps to $8.3B valuation, TechCrunch, July 1, 2026
• Fireworks AI, Fireworks Series D: $1.5B Raised at $17.5B Valuation, July 15, 2026
• Dominic-Madori Davis, AI inference startup Baseten reportedly raising $1.5B months after its last mega-round, TechCrunch, June 18, 2026
• Baseten, Announcing our Series F, June 22, 2026
• Rebecca Torrence and Bailey Lipschultz, Baseten Set for $30 Billion Valuation With New Funding, Bloomberg, October 30, 2026
• Modal, Modal's Series C: Raising $355M at a $4.65B valuation, May 21, 2026
• Modal, Modal crosses $1B in Annualized Run Rate and raises $750M Series D, October 14, 2026
• Rich Wong and Gonzalo Mocorrea, Our investment in Modal, Accel, October 14, 2026
• Databricks, Databricks Agrees to Acquire Neon to Deliver Serverless Postgres for Developers + AI Agents, May 14, 2025
• Snowflake, Snowflake Acquires Crunchy Data to Bring Enterprise Ready Postgres Offering to the AI Data Cloud, Business Wire, June 2, 2025
• ClickHouse, ClickHouse welcomes Langfuse: The future of open-source LLM observability, January 16, 2026
• Dynatrace, Dynatrace to Acquire AI Observability Leader Arize, August 13, 2026
• Dynatrace, Dynatrace Completes Acquisition of Arize, October 2026
• Marina Temkin, Inference startup Inferact lands $150M to commercialize vLLM, TechCrunch, January 22, 2026
• RadixArk, Our $100M Seed to Build Open Infrastructure for Frontier AI, May 5, 2026
• James Thomason, Exclusive: Browserbase launches headless browser platform that lets LLMs automate web tasks, VentureBeat, June 6, 2024

Salient PR works with AI infrastructure companies including Tenstorrent and Browserbase. Funding, valuation and revenue figures reflect public reporting as of November 2026.