AI & Tech
2026 AI Infrastructure Investment Map — What $297B in Record VC Spending Reveals
Published: 2026-05-11
Q1 2026 global venture funding reached $297B — a record for any single quarter. 81%, or roughly $240B, flowed into AI. The number is impressive, but the direction of capital tells the more important story for founders.
From Training to Inference — Where Capital Is Moving
The 2023–2024 cycle was defined by massive foundation model training investments. OpenAI, Anthropic, xAI raised billions to build and scale models. Q1 2026 looks different. Roughly two-thirds of AI capex is now allocated to inference — running models, not training them.
The logic is straightforward. Models are now powerful enough. The bottleneck isn’t smarter models; it’s faster, cheaper model execution at scale.
Among 25 AI infrastructure deals totaling $10.64B over the past year, inference platforms and inference chips represent the largest share:
| Company | Amount | Segment |
|---|---|---|
| Cerebras Systems (IPO) | $3.5B raise / $26.6B valuation | AI chips (WSE-3) |
| ElevenLabs | $500M Series D ($11B) | Voice AI |
| Nexthop AI | $500M | AI networking |
| Runway | $315M Series E ($5.3B) | Media generation |
| Baseten | $300M Series E ($5B) | Inference platform |
| Sierra | $950M ($15B+) | AI agents |
| RadixArk (SGLang) | $100M Seed ($400M) | Inference engine |
| Inferact | $150M Seed ($800M) | AI inference |
Open Source Is Raising Venture Rounds
The most structurally interesting signal: RadixArk. The team behind SGLang — an open-source inference serving engine that rivals vLLM in adoption — raised a $100M Seed round led by Accel and Spark Capital at a $400M valuation.
The playbook is familiar from the database era (MongoDB, HashiCorp, Elastic): win the ecosystem with open source, monetize through managed cloud services, enterprise support, and hosted platforms. What’s different is the timeline. In ML infrastructure, open source can become the de facto standard within six months, and the VC round follows the month after.
NVIDIA Alternatives Are Going Institutional
Cerebras’s WSE-3 IPO at a $26.6B valuation signals that NVIDIA alternatives have crossed the threshold from “interesting experiment” to “institutional-grade investment thesis.” Groq’s LPU expansion, SambaNova, Nexthop AI’s networking layer — capital is flowing into every part of the stack that can reduce NVIDIA dependency.
This matters for procurement teams as much as investors. Competition at the chip and networking layer will change pricing power dynamics within 2–3 years.
Four Signals Founders Should Act On
1. Every layer of the inference stack is becoming a separate company. Inference compiler (SGLang), inference chip (Cerebras), inference platform (Baseten), inference networking (Nexthop) — each layer is raising independent rounds. Owning one layer well is sufficient.
2. Open source to $400M valuation in under a year is now a documented path. RadixArk proves it. Build the ecosystem first; monetize second.
3. AI agents are infrastructure, not SaaS. Sierra’s positioning at $15B+ valuation isn’t selling an AI tool — it’s selling the platform that enterprise customers build their own agents on top of. The platform layer wins on distribution, not features.
4. Capital concentration creates an infrastructure demand floor. The top three deals (OpenAI, Anthropic, xAI) absorbed 45.8% of total AI capital. Every dollar those companies spend on compute is revenue for infrastructure founders. The demand base is as certain as any in venture history.
Sources
Related Content