Robin Bordoli

Post 9/10 AI Prediction #9 for 2026: 𝐀𝐧𝐭𝐡𝐫𝐨𝐩𝐢𝐜 𝐈𝐏𝐎 𝐲𝐞𝐬, 𝐎𝐩𝐞𝐧𝐀𝐈 𝐈𝐏𝐎 𝐧𝐨 Capital markets matter. In 2025, there were ~200 IPOs and only two AI IPOs, raising roughly $2B combined. That will be dwarfed in 2026. I expect Anthropic to IPO and raise at least $30B in equity. Despite being larger, OpenAI is further from profitability and likely to face more turbulence in public markets given its less clear path to profitability. I predict OpenAI will choose to stay private longer, supported by alternative capital sources, and unparalleled consumer reach with their 800 million weekly active users of ChatGPT that processes 2.5 billion prompts daily. Anthropic’s IPO valuation would likely be in the $300 to $350B range, based on 2025 $9B revenue and 2026 $20-26B revenue drive by enterprise adoption of coding agents powered by Claude. This level of valuation would be a seminal moment for AI adoption as the largest ever tech IPO.

Post 10/10 AI Prediction #10 for 2026: 𝐈𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 𝐞𝐜𝐨𝐧𝐨𝐦𝐢𝐜𝐬 𝐨𝐯𝐞𝐫𝐭𝐚𝐤𝐞 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐞𝐜𝐨𝐧𝐨𝐦𝐢𝐜𝐬 For the last few years, AI discourse has been dominated by training runs, parameter counts, and capex arms races. In 2026, the center of gravity decisively shifts to inference. Signals are everywhere: from NVIDIA’s decisive acquihire and licensing of Groq for $20B to the rapid growth of inference providers like Fireworks AI, Together AI, and Baseten. These signals reflect where value is moving. Not just who can train the biggest model, but who can serve intelligence cheapest, fastest, and at scale. Fleets of always-on agents, multimodal inference, long-context reasoning, and real-time systems have different requirements than multi-month-long training runs. This reshapes everything: • Model architectures optimize for latency, memory, and cost-per-task • Smaller, specialized, and domain-tuned models win in production • Inference scaffolding - routing, batching, caching, and model selection -matters as much as raw model quality I predict we will see: * accelerated proprietary silicon roadmaps from the hyperscalers * split-brain infrastructure usage: GPUs for training, ASICs for inference * stronger benchmarking discipline: $ / token / p99 latency across vendors * latency-first SKUs: new model variants optimized for real-time agents