Madrona’s annual IA40 list shows an AI industry splitting in two
THE SO WHAT
Madrona’s IA40 showing 45 private AI companies with $410B raised—and 92% of that concentrated in Anthropic, OpenAI, and Databricks—confirms a barbell market: a few capital magnets and a long tail. If you’re not in the capital gravity well, your edge has to be distribution, domain depth, or unit economics, not “we’ll raise our way into scale.”
READ THE SOURCE
MORE FROM THE WIRE
Startups & VentureAfterQuery reportedly becomes Y Combinator’s fastest-ever unicorn, now valued at $3.2B
A jump from a $300M valuation to $3.2B in five months for an AI model-training startup tells you where late-stage capital thinks the margin pool is—custom training and infra, not just apps. If you’re building on top of these stacks, assume your upstream providers will move aggressively into higher-value services over the next 12–24 months.
Startups & VentureSources: AfterQuery, which sells coding and finance training data to AI labs, has hit a valuation of $3.2B, up from $300M in April, and is profitable
A jump from $300 million to $3.2 billion in months for a training-data vendor shows that high-quality, rights-cleared corpora are becoming a core economic choke point. If you own differentiated data — especially code or finance — you’re no longer just a SaaS or services company, you’re in the model supply chain.
Startups & VentureThrive’s Kushner defends involvement in FIFA mess, hires Elon’s go-to lawyer
When a top-tier venture firm becomes part of a global sports governance scandal, LP and regulator scrutiny on side deals and influence networks goes up across the asset class. If you’re running a fund or late-stage startup, assume reputational due diligence on your cap table and advisory relationships is tightening, not loosening.
Startups & VentureWafer, which makes AI agents that optimize open-source models for a business's workload, raised a $40M Series A, a source says at a $200M+ valuation
A $40M Series A at a $200M+ valuation for Wafer’s AI agents that tune open-source models on non-NVIDIA chips is a bet that inference optimization itself is a profit center, not just a feature. If you’re building on open models, assume your cost and latency advantage will be competed away—start benchmarking against specialized optimizers now.