The era of buying raw, unmediated model access is no longer the default enterprise play. As buyers confront unpredictable inference bills, data-governance risk, and a growing sprawl of model vendors, the capital is moving toward infrastructure that can route, secure, compress, and control AI workloads rather than simply expose another API.
According to LeadPrysm’s proprietary tracking, we have mapped 244 AI startup raises in just the last 30 days across 27 countries. Vertical SaaS AI remains the most active sub-vertical with 63 deals, while AI Infrastructure (30 deals) and AI Agents (24 deals) continue to attract the most strategic capital. The throughline is clear: the bottleneck has shifted from raw model quality to governance, orchestration, sovereignty, and cost control.
The Shift to Local and Sovereign Compute Economics
For much of the last two years, the operating assumption was that enterprise AI would run through centralized cloud APIs. The newest funding wave suggests a sharp reversal toward local execution, sovereign deployments, and tighter control over data movement.
Ollama is a clean example. The open-source developer tool and platform raised a $65 million Series B led by Theory Ventures, with participation from Benchmark, 8VC, and Y Combinator. TechCrunch reported that Ollama says it now serves 8.9 million monthly active developers, underscoring how quickly local-first model execution has become a mainstream developer workflow. (techcrunch.com)
That same sovereignty theme is showing up in regulated markets:
- AVELIN AI raised $3.7 million in pre-seed funding from angel investors to scale its sovereign AI platform for enterprises and governments. AVELIN says it focuses on “Cross-Model Fusion” and on-premise deployment for regulated environments. (tradingview.com)
- Sherpa.ai recently announced new funding led by Forgepoint Capital, with participation from Mundi Ventures, Ekarpen, Allegra Holdings, and SETT. Public reporting is inconsistent on the headline amount, so the safer takeaway is the investor mix and the company’s focus: privacy-preserving federated learning for collaborative AI on private datasets. (cincodias.elpais.com)
As enterprises realize that data privacy is not optional, AI dev tools are becoming trust layers, not productivity widgets.
Managing the Multi-Agent Mess: Governance and Orchestration
The next bottleneck is no longer whether a model can answer a question. It is whether a stack of models and agents can act safely, consistently, and within policy.
Lyzr is a good example of how that demand is being framed in the market. Bloomberg reported that Lyzr AI closed a $14.5 million funding round led by Accenture, with Rocketship VC participating, and that the company’s valuation rose to $250 million. Lyzr is positioning itself around enterprise agent infrastructure: orchestration, governance, and the ability to run collaborative AI agents in controlled environments. (bloomberg.com)
That same governance problem is starting to define the AI security market. If agents can call tools, move data, and trigger workflows, then the buyer’s concern shifts from “Is the model accurate?” to “What can this system do, and under whose control?” That is also why products like AIsa, which says it is building a transaction network and programmable interface for AI agents, are getting attention. AIsa announced $6.5 million in total funding to date, including a new seed round co-led by Alibaba and Tribe Capital. (globenewswire.com)
Hardware Heterogeneity and Inference Optimization
The infrastructure wave is also being driven by the need to make inference cheaper, faster, and more deployable across heterogeneous hardware.
SambaNova is still one of the most visible examples. The company announced the first close of a $1 billion Series F, valuing it at $11 billion post-money. The deal was led by General Atlantic, and SambaNova also said JPMorganChase selected it as an inference infrastructure partner for secure on-prem AI inference. (sambanova.ai)
Other infrastructure rounds reinforcing the same direction include:
- Prime Intellect announced a $130 million Series A to build what it calls the open superintelligence stack. The company says it is focused on decentralized training, RL infrastructure, and distributed compute. (primeintellect.ai)
- Bespoke Labs raised $40 million in Series A financing led by Wing VC, with participation from Mayfield and others, to build environments for training and evaluating agents. (streetinsider.com)
- Gradium extended its seed financing to $100 million and said NVIDIA is joining as a new investor. The Paris-based company, spun out of Kyutai, is building low-latency audio models for real-time voice agents. (gradium.ai)
- Ora Computing describes itself as a model-compression startup focused on making AI models smaller and more efficient for deployment. Its public materials say it raised a €3.5 million seed round. (oracomputing.com)
What This Means for Startups Selling to the AI Ecosystem
If you are building tools for AI startups or enterprise buyers, the playbook is changing fast:
- Sell predictability, not just performance. Buyers care about cost control, routing, and observability as much as raw model output.
- Design for hybrid and on-prem deployments. Ollama, AVELIN AI, and SambaNova all point to the same requirement: local control is becoming a feature, not a niche preference. (techcrunch.com)
- Treat orchestration as a security problem. The rise of Lyzr and AIsa suggests that the AI stack is becoming a control-plane business, not just a model-access business.