AI Infrastructure: what are the top startups now?

Which AI cloud startups are winning the GPU land grab?

CoreWeave is still the clear leader; Lambda and Crusoe are the most serious private challengers; Nscale and TensorWave are the newer names that suddenly deserve attention.

CoreWeave is ahead because its numbers are in another category. In May 2026, it reported revenue backlog near $100 billion and more than 1 GW of active power, with a stated path toward more than 8 GW by 2030. That gives us a hard comparison point: CoreWeave is not just raising money to build capacity. Actually, it already has massive contracted demand and physical power capacity behind it.

Lambda is the closest “classic GPU cloud” challenger. Its November 2025 Series E brought in more than $1.5 billion, and the company framed the round around gigawatt-scale AI factories. Compared with CoreWeave, Lambda is less transparent on backlog and revenue because it is still private. But it has something very valuable: strategic demand from hyperscalers, enterprises, and frontier labs that do not want to depend only on AWS, Azure, Google, or CoreWeave.

Crusoe is not trying to beat CoreWeave on cloud software alone. Its edge is energy and data-center execution. The company raised $1.375 billion in October 2025 at a valuation above $10 billion, and its pitch is very direct: the bottleneck in AI infrastructure is not only GPUs, it is power. That makes Crusoe more interesting than a generic GPU reseller. Today, if you can secure power and build AI factories faster, you are part of the compute stack.

Nscale is the fastest-rising European AI cloud name. It raised $2 billion in March 2026 at a $14.6 billion valuation, after already raising $1.1 billion in September 2025. That pace matters. Compared with Lambda and Crusoe, Nscale has less long-term proof, but it has the strongest recent “sovereign AI infrastructure” signal. Europe wants regional compute capacity, and Nscale is one of the few startups now funded at a scale that matches that ambition.

TensorWave is earlier, but it is the most interesting contrarian compute bet. Its June 2026 $350 million raise valued it at $1.55 billion, and the company is building around AMD chips rather than Nvidia. That is the whole point. TensorWave is not yet in CoreWeave’s league, but it gives buyers a second supply chain. In a market where everyone is fighting for Nvidia capacity, that alone makes it worth watching.

So, CoreWeave is the benchmark, Lambda and Crusoe are the credible private challengers, Nscale is the emerging sovereign-infrastructure bet, and TensorWave is the AMD alternative that becomes more important if Nvidia scarcity stays painful.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Which inference startups are turning AI usage into real revenue?

Fireworks AI, Together AI, Baseten, and Modal are the leading pack. Fireworks and Together look strongest on scale; Baseten is better positioned for enterprise deployment; Modal is the fast-rising developer infrastructure name.

This category is where the market feels hottest now. Training runs are huge, but they are episodic. Inference is the daily business. Every chatbot response, coding agent, search assistant, voice workflow, and AI feature creates ongoing compute demand. That is why inference startups are suddenly being valued like core infrastructure companies.

Fireworks AI looks like the most explosive pure inference company right now. Recent market estimates put it at roughly $800 million in annualized revenue in May 2026, up from about $305 million at the end of 2025. Even if we treat those estimates carefully because they are private-company numbers, the direction is hard to ignore. Fireworks appears to be scaling faster than most model-serving startups because it sits close to production traffic, not experiments.

Together AI is broader. It combines open-model hosting, fine-tuning, training, and inference. In March 2026, reports said Together was in talks to raise around $1 billion at a $7.5 billion valuation, more than double its prior valuation less than 14 months earlier. Compared with Fireworks, Together looks less like a narrow inference API and more like a full open-model cloud.

Baseten is the enterprise deployment winner. Its January 2026 $300 million Series E valued it at $5 billion, only three months after its previous round. Nvidia’s participation makes the signal stronger because Nvidia has been backing the companies that help turn chips into usable production systems. Baseten is not trying to be the cheapest token API. It is closer to the layer enterprises need when they deploy custom models, manage scale, and care about reliability.

Modal is moving fast because it solves a different problem. Its May 2026 $355 million Series C valued it at $4.65 billion, more than four times its September 2025 valuation. That kind of jump tells us developers want serverless GPU infrastructure, not another complicated cloud console. Modal is probably not ahead of Fireworks or Together on raw inference volume, but it may be closer to the developer workflow where new AI apps are born.

The ranking depends on what we measure. For usage scale, Fireworks and Together look strongest. For enterprise model deployment, Baseten is cleaner. For developer-native GPU workflows, Modal is the one moving fastest lately.

Which AI chip startups still look credible against Nvidia?

Cerebras is the most validated; Groq has the strongest strategic validation; Etched is the most exciting new hardware bet; Tenstorrent and MatX stay on the watchlist with thinner recent proof.

Nobody is “beating Nvidia” across the board. That is the wrong question. The useful question is whether any startup can win a specific slice of the AI hardware stack: wafer-scale systems, ultra-fast inference, transformer-specialized chips, open architectures, or sovereign alternatives.

Cerebras is the most proven because it has public-market validation and real revenue. Its May 2026 IPO put it in a different evidence bucket from private chip startups. Recent IPO analysis pointed to about $510 million of 2025 revenue and a major OpenAI-linked contract pipeline. That does not mean Cerebras has solved the whole Nvidia problem, but it does mean it has crossed from “interesting chip architecture” into “commercial AI infrastructure company.”

Groq is the strangest case because Nvidia’s December 2025 licensing and talent deal changed the story. Groq stayed independent, but Nvidia licensed its inference technology and hired key leaders. Compared with Cerebras, Groq has less clean standalone momentum now because part of the team moved. Still, the validation is powerful. Nvidia does not do that kind of deal unless it sees something important in Groq’s inference architecture.

Etched is the highest-upside emerging name. In January 2026, it reportedly raised about $500 million at a $5 billion valuation to build transformer-specialized chips. The risk is obvious: it still needs to prove production, software, customers, and supply chain. But the bet is sharp. If a huge share of AI inference remains transformer-like, a chip designed around that assumption could be much more efficient than a general GPU for certain workloads.

Tenstorrent is still credible, but the story is less hot right now. Its December 2024 round raised about $693 million at a roughly $2.6 billion valuation, backed by major strategic investors. The company matters because RISC-V and alternative accelerator designs are useful for countries and companies that want more control. But compared with Etched, the recent market signal is less intense.

MatX is earlier. It has credible technical DNA from ex-Google TPU talent and raised $80 million in January 2025, but it is not yet in the same proof tier as Cerebras, Groq, or Etched. We would call it a watchlist name, not a leader.

So the clean ranking today is Cerebras first on commercial proof, Groq first on strategic validation, Etched first on emerging upside, then Tenstorrent and MatX as longer-cycle hardware bets.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Which startups are solving the data-movement bottleneck?

Ayar Labs is the clearest leader; Celestial AI and Enfabrica are serious challengers; Lightmatter remains important but needs fresher public proof.

This category matters more than most people realize. AI clusters do not only need faster chips but chips to actually talk to each other without wasting huge amounts of time and energy. These days, moving data can be as painful as computing it.

Ayar Labs is the strongest current name because it has the freshest hard signal. In March 2026, it closed a $500 million Series E at a $3.75 billion valuation, bringing total funding to $870 million. More importantly, the company said the money would go toward high-volume production and test capacity for co-packaged optics. That puts Ayar closer to scaling a real infrastructure component, not just pitching photonics as a future idea.

Celestial AI is the next important name. It raised $250 million in March 2025 for its Photonic Fabric technology, and it sits in the same core problem area: reducing the cost and latency of moving data across AI systems. Compared with Ayar, Celestial has a slightly older signal in this snapshot, but the market logic is similar. Optical connectivity is becoming part of AI scale-up.

Enfabrica attacks the problem from AI networking silicon. Its value is easier to understand if we compare it with Ayar. Ayar is about optical I/O and co-packaged optics. Enfabrica is more about the network interface and fabric that connect huge accelerator clusters. If AI data centers keep growing, both layers matter.

Lightmatter is still worth watching because photonic computing and interconnects remain strategically relevant. But we should be honest: based on the freshest available signals, Ayar has the cleaner current leadership claim.

Ayar is the top name now, Celestial and Enfabrica are credible challengers, and Lightmatter stays in the category because the technical direction still matters.

Which AI data startups gained power after the Scale AI shock?

Scale AI is still the leader by strategic value, but Surge AI and Mercor are the more interesting “what changed lately?” names. Turing remains relevant, though less hot.

This market changed after Meta’s Scale AI investment. In June 2025, Meta invested about $14.3 billion for a 49% stake in Scale AI, implying a valuation above $29 billion. That proved something very clearly: training data is not a back-office labeling service anymore. It is strategic AI infrastructure.

But Scale’s strength also created an opening for competitors. If you are OpenAI, Google, Anthropic, or another frontier lab, do you want your sensitive training workflows close to Meta? Even if Scale remains independent operationally, the perception problem is real. That is why neutral data providers matter more now.

Surge AI is the clearest beneficiary. Reports in 2025 said it had reached more than $1 billion in revenue and was exploring a raise at a valuation above $15 billion. Compared with Scale, Surge has less strategic-brand gravity, but it may now have a cleaner neutrality story. In this category, neutrality is not a soft factor. It can decide where frontier labs send their most sensitive data work.

Mercor is the fastest-rising expert-data marketplace. In October 2025, reports said it was close to raising $350 million at a $10 billion valuation. Its role is different from classic labeling. Mercor connects AI labs with domain experts who can create, evaluate, and improve training data. That becomes more valuable as models move from generic internet knowledge toward reasoning, coding, finance, law, science, and other expert domains.

Turing is the more mature operator. It raised $111 million in March 2025 at a $2.2 billion valuation, and earlier reporting pointed to around $300 million in revenue after rapid growth. Compared with Mercor, it looks less explosive. Compared with many smaller data startups, it has more operating proof.

The ranking has changed. Scale is still the giant, but its Meta link makes the category less settled. Surge is the strongest neutral-scale challenger, Mercor is the most exciting expert-data upstart, and Turing is the reliable enterprise operator.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Which vector database startups are still worth caring about?

Pinecone, Weaviate, Qdrant, Zilliz, and Chroma still matter, but this is no longer the hottest part of AI infrastructure.

This is one place where we should not force excitement. Vector databases were one of the clearest AI infrastructure categories in 2023 and 2024. Today, the market is more complicated. Retrieval is still essential, but the standalone vector database story has cooled because vector search is becoming a feature inside broader databases, cloud platforms, and AI frameworks.

Pinecone remains the most recognizable managed vector database company. It still has the clearest enterprise brand in the category. If a buyer wants a specialized hosted vector database, Pinecone is usually one of the first names considered.

Weaviate and Qdrant are stronger when the buyer cares about open source, deployment control, and cost. Compared with Pinecone, they feel less like pure managed infrastructure and more like flexible building blocks. That matters for teams that want retrieval inside their own stack.

Zilliz remains important because of Milvus. Its strength is large-scale vector infrastructure linked to an open-source project with real adoption. Chroma is more developer-friendly and culturally visible, but we would be careful about ranking it too high without fresher commercial evidence.

So the answer here is intentionally less dramatic. Pinecone is still the default managed leader. Weaviate, Qdrant, and Zilliz have durable open-source or scale arguments. Chroma is worth watching. But the freshest infrastructure heat has moved toward inference, agents, evaluation, and context engineering.

Which startups are becoming the rails for AI agents?

LangChain is the default leader; LlamaIndex is strongest around data-to-agent workflows; CopilotKit and Composio are the fresher emerging names; Temporal is not AI-native, but agents make it newly relevant.

Agent infrastructure is noisy. A lot of companies say they are building “agent rails,” but many only have demos, GitHub stars, or vague developer interest. So we should rank this category by what problem each company actually owns.

LangChain is the leader because it expanded beyond an open-source framework. In October 2025, it raised $125 million at a $1.25 billion valuation. The important part is the product expansion: LangGraph for agent orchestration and LangSmith for tracing and evaluation. That gives LangChain a broader control point than a simple library.

LlamaIndex is strongest when the agent problem starts with data. If a company needs to connect documents, databases, retrieval pipelines, and workflows into an agentic system, LlamaIndex has a clear reason to exist. Compared with LangChain, it feels less dominant as a general agent platform, but more focused on the data-context layer.

CopilotKit is the emerging app-native agent name. Its May 2026 $27 million raise is not huge compared with LangChain, but the positioning is timely. Agents increasingly need to live inside existing products, not in separate chat windows. CopilotKit is interesting because it starts from that product-embedding problem.

Composio is worth watching because agents need tools, auth, API actions, and permissioning. That sounds boring, but it is exactly where prototypes break. Compared with LangChain, Composio is narrower. That can be a weakness, but also a strength if tool-use infrastructure becomes a must-have layer.

Temporal is not an AI startup in the narrow sense, but it belongs in the conversation. Long-running agents need durable workflows. They need retries, state, scheduling, and failure handling. That is Temporal’s world. The AI wave did not create the company, but it makes the need more obvious.

LangChain is the agent-infrastructure default, LlamaIndex owns a strong data-context wedge, CopilotKit and Composio are the emerging application and tool-use plays, and Temporal is the older workflow layer that suddenly fits the agent era.

Which observability and evaluation startups are becoming must-have?

LangSmith, Arize/Phoenix, Braintrust, Langfuse, Galileo, and Helicone stand out, but no single company owns the category yet.

The reason this category matters now is that companies are discovering a painful truth: shipping AI is easy, keeping it reliable is hard. Prompts change. Models change. Costs move. Latency spikes. Agents fail in strange ways. That creates a real need for tracing, evals, regression tests, and monitoring.

LangSmith has the strongest distribution advantage because it is tied to LangChain and LangGraph. If a team already builds agents there, LangSmith is the natural place to debug and evaluate them. That gives it a stronger ecosystem pull than most standalone eval tools.

Arize has the strongest old-world ML observability background. Phoenix gives it an open-source wedge, while Arize’s enterprise product speaks to teams that already understand drift, monitoring, and production ML. Compared with LangSmith, Arize is less tied to one agent framework. That can be useful for larger companies with mixed stacks.

Braintrust is probably the cleanest eval-native startup. Its strength is treating AI quality like an engineering loop: datasets, experiments, scoring, human review, and regression testing. Compared with generic observability tools, Braintrust is more focused on whether the model or agent actually got better.

Langfuse is important because of open-source and self-hosting. Many teams do not want to send prompts, traces, or customer data into a closed third-party system. Langfuse wins when control matters.

Galileo and Helicone are useful in narrower ways. Galileo leans toward enterprise evaluation and monitoring. Helicone is simple and developer-friendly for logging, cost, and usage tracking. Neither clearly dominates, but both solve real production problems.

Today, this category is still fragmented. LangSmith has ecosystem pull, Arize has enterprise observability credibility, Braintrust has the clearest eval focus, and Langfuse has the open-source/self-hosted argument.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Which AI security startups are suddenly becoming urgent?

Noma Security is the strongest independent name now; Lakera, Protect AI, Prompt Security, and Pangea were validated by acquisition; HiddenLayer remains a credible model-security specialist.

AI security is one of the clearest “now it matters” categories. When AI was mostly employees using chatbots, the risk was data leakage and policy control. With agents, the risk gets bigger. Agents can touch tools, execute actions, access internal data, and make mistakes at machine speed.

Noma Security is the leading independent startup to watch. In July 2025, it raised $100 million in Series B funding less than a year after its prior round. Its focus on AI and agent security fits where the market is going. Compared with older AI-security companies, Noma feels more native to the agent era.

The tricky part is that several strong names have already been bought. Protect AI was acquired by Palo Alto Networks in July 2025. Prompt Security was acquired by SentinelOne in August 2025. Lakera was acquired by Check Point in 2025. Pangea was acquired by CrowdStrike in 2025. That acquisition wave tells us the category is real, but it also makes the independent startup leaderboard unstable.

HiddenLayer remains relevant because model threat detection and AI attack surfaces are not going away. It is less visibly hot than Noma in the most recent funding cycle, but the problem it attacks is still central.

Which model-routing startups are becoming hidden control points?

OpenRouter is the clearest independent gateway; Together AI and Fireworks AI matter because they combine inference and distribution; Replicate and Fal remain important for model access, especially outside plain text LLMs.

This category is easy to underestimate. Developers do not want to rebuild their app every time a new model gets cheaper, faster, or better. Instead, they want routing, fallback, price comparison, model access, usage controls, and one API layer. That is why gateways can become control points.

OpenRouter is the cleanest independent startup in this lane. Its value is neutrality. It actually does not need to own the best model or the biggest GPU cluster. But it does need to sit between developers and many models, then make switching easy. In a market where model rankings change constantly, that is a real infrastructure position.

Together AI and Fireworks AI appear again here for a different reason. As seen above, they are strong inference platforms. But they also distribute access to open and third-party models. That gives them a second advantage: developers can build around their API surface, not just their raw compute.

Replicate remains important for model discovery and deployment, especially for open-source and multimodal models. Its independent startup status changed after acquisition, but its role shows why this layer matters. Easy access to community models became valuable enough for larger platforms to absorb.

Fal is the more specific media-inference play. It is not trying to route every LLM. Instead, we should look at it as infrastructure for image, video, and generative media workloads where latency and GPU scheduling matter.

So who are the top AI infrastructure startups right now?

The current top tier is CoreWeave, Fireworks AI, Together AI, Lambda, Crusoe, Baseten, Modal, Scale AI, Surge AI, Mercor, Cerebras, Groq, Etched, Ayar Labs, LangChain, Noma Security, and OpenRouter.

CoreWeave sits alone at the top of the compute layer because its backlog and power footprint are now measurable at a scale other startups cannot match. Lambda and Crusoe are the strongest private compute challengers, but for different reasons. Lambda is closer to the GPU-cloud demand story. Crusoe is closer to the power-and-data-center bottleneck.

Fireworks AI and Together AI are the strongest inference names because they sit where usage becomes recurring revenue. Baseten and Modal are slightly different: Baseten looks strongest for enterprise model deployment, while Modal is the developer-native GPU layer moving fastest lately.

Scale AI is still the strategic data giant, but Surge AI and Mercor are where the fresh story is. Surge has the revenue-scale and neutrality argument. Mercor has the expert-data argument. That distinction matters because frontier AI labs now need specialized human knowledge, not just generic labeling.

Cerebras, Groq, and Etched are the most important chip names, but their proof points are different. Cerebras has public-market and revenue validation. Groq has Nvidia validation. Etched has the strongest emerging architectural bet. Ayar Labs belongs beside them because the scaling bottleneck increasingly includes interconnects, not just chips.

LangChain, Noma Security, and OpenRouter round out the list because infrastructure is moving up the stack. Agents need orchestration. Agents need security. Apps need model routing. These categories are not as financially mature as compute or inference, but they are where developer behavior is changing now.

AI infrastructure now has several control points. CoreWeave controls scarce compute. Fireworks and Together control production inference. Scale, Surge, and Mercor control data supply. Cerebras, Groq, Etched, and Ayar Labs attack the hardware bottleneck. LangChain, Noma, and OpenRouter sit closer to the new software layer that developers actually touch.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Category	Startups selected and why
GPU cloud and neocloud	CoreWeave leads on backlog and power scale; Lambda is the strongest classic GPU-cloud challenger; Crusoe is strongest on power-first AI factories; Nscale is the fast-rising sovereign compute bet; TensorWave is the AMD alternative to watch
Inference platforms	Fireworks AI looks strongest on recent revenue acceleration; Together AI is broader across open-model cloud, training, and inference; Baseten is strongest for enterprise deployment; Modal is the fastest-rising developer GPU workflow layer
AI chips	Cerebras leads on commercial and public-market proof; Groq has the strongest strategic validation from Nvidia; Etched is the highest-upside new architecture bet; Tenstorrent and MatX remain credible but less proven
Interconnects and networking	Ayar Labs has the clearest current leadership signal; Celestial AI and Enfabrica are serious challengers; Lightmatter remains important but needs fresher public proof
Training and expert data	Scale AI is still the strategic data giant; Surge AI is the strongest neutral-scale challenger; Mercor is the fastest-rising expert-data marketplace; Turing is the mature operator
Vector databases and AI memory	Pinecone remains the managed leader; Weaviate, Qdrant, and Zilliz have stronger open-source or scale arguments; Chroma stays visible with developers, but the category is cooler now
Agent infrastructure	LangChain is the default agent-infrastructure leader; LlamaIndex owns the data-context wedge; CopilotKit and Composio are emerging around app-native agents and tool use; Temporal matters for durable workflows
Observability and evals	LangSmith wins on ecosystem pull; Arize/Phoenix wins on ML observability depth; Braintrust is strongest on eval workflows; Langfuse is strongest for open-source/self-hosted teams
AI security	Noma Security is the top independent name; Protect AI, Prompt Security, Lakera, and Pangea prove the category through acquisitions; HiddenLayer remains a credible model-security specialist
Model routing and gateways	OpenRouter is the clearest neutral gateway; Together AI and Fireworks AI combine inference with distribution; Replicate validated model catalogs; Fal is the stronger media-inference specialist

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.