The State of Tech: AI x VC, Q1 2026

The state of play: where tech, AI, and capital stand right now.

I’m sharing our full LP presentation covering AI and the VC market in S2 2025. We believe in open-sourcing knowledge: producing rigorous data and insights is how you strengthen a tech ecosystem. It builds shared understanding and attracts capital.

Whether you’re a founder, an investor, or building in the space, we hope it proves useful: Galion.exe, The State of Tech, Q1 2026

Below, the key takeaways.

1.1. AI: the industrialisation paradox

The capability surge continues, but deployment lags. The frontier is a triopoly, and verification sets the real pace of adoption.

Compute is becoming a geopolitical asset. Hyperscaler capex is projected to exceed $200bn in 2025. Only five to seven players globally can compete at frontier scale; everyone else must find defensible niches. China trails the US by seven or more months on average; export controls appear to be working. Meanwhile, available compute doubles every nine months, but demand growth outpaces supply. NVIDIA concentration (60%+ of AI compute) creates systemic supply chain risk, though alternative architectures are gaining traction.

Capability accelerated 2x in 2024, not just from more FLOPs, but from new paradigms: reinforcement learning, data quality improvements, and test-time compute (spending more FLOPs at inference, not training). Pretraining is hitting diminishing returns. At current slope, ECI 160+ by end of 2026, unlocking serious agentic and scientific automation potential. For more on how capabilities translate (or fail to translate) into defensible markets, see Surviving the Reset.

Verification, not generation, determines adoption speed. Domains with clear reward signals (code, maths, formal reasoning) improve rapidly via RL. Domains with ambiguous success criteria lag 12 to 24 months. This explains why progress feels jagged and why strategic advantage lies in measurement, not just cognition. Intelligence Isn’t the Moat Anymore expands on this asymmetry.

The frontier is a triopoly. Google, OpenAI, and Anthropic control base models. Meta has exited the frontier race (for now). Open source trails by six to nine months. Since GPT-4, every frontier model has come from US labs.

Distribution beats raw performance. Gemini’s growth (+240% YoY) is Android integration, not model superiority. Only 9% of consumers pay for multiple AI subscriptions. ChatGPT still dominates consumer (800M+ WAU), but Anthropic flipped enterprise: 40% share (up from 12% in 2023). OpenAI hit $20bn ARR, 2.5x Anthropic, but Anthropic’s growth from $1bn to $9bn ARR in 18 months shows enterprise is contestable.

AI tourists bias all traction metrics. Users pay $20-200 to test the latest AI release, then churn within 90 days. 4o image generation added 1M users/hour at peak; most were gone within weeks. Only 9% of consumers pay for multiple AI subscriptions simultaneously.

Agents are production-ready at ~L2. The question shifts from “can it do the task?” to “can it do it reliably, safely, and affordably?” MCP (Model Context Protocol) is emerging as the “USB-C for agents”, enabling universal tool connectivity across models and platforms. Reliable enough is the operative standard: like aviation or medicine, the goal is not error elimination but error management at acceptable rates.

The deployment paradox persists. Model capabilities advance faster than enterprise ability to absorb. Bottlenecks are shifting from technology to change management, data infrastructure, and governance frameworks. Winners will be those who solve the “last mile”: integration, trust, reliability. For more on how constraints shape deployment, see The Deployment Paradox and Deployment Metabolism.

1.2. Global VC: the bifurcated recovery

Recovery is real but AI-driven. Strip out AI mega-rounds and funding is essentially flat. The middle is hollowing out.

Global VC surged to $512bn (+28% YoY). But deal count remains depressed (~8,000 per quarter vs 12,000+ in 2021): fewer deals, bigger cheques. AI captured 64%+ of all VC deployed (vs ~39% in 2023), reaching $211bn, up 85% YoY.

Capital concentration at scale. The top five companies (OpenAI, Anthropic, xAI, SpaceX, Databricks) raised $84bn: 20% of all 2025 VC. 30%+ of quarterly funding now flows to $500m+ rounds; 68 companies raised $500m+ in 2025.

US dominance widens. $339bn (+44% YoY), capturing 64% of global VC. The San Francisco Bay Area alone raised $122bn. Europe remains resilient at €44bn (+7% YoY), the highest since 2022, but with declining global share (11-17% vs 16% in 2024). Asia at multi-year lows.

GP fundraising at seven-year low. $118bn raised, down 46% YoY. Only ~1,000 funds closed vs 4,500+ in 2022. Top 10 funds raised more than all others combined. a16z raised 18% of all US VC in 2025: $15bn across six funds, now $90bn+ AUM.

The liquidity problem persists. Most LPs are freezing new commitments, waiting for distributions from older funds. Secondaries now account for 71% of liquidity value. IPO window remains half-open: 21 tech IPOs in 2025 (+162% YoY), but still far below 2020-2021 levels. M&A volume hit record: $587bn across 326 deals (>$50m).

Deeptech, TechBio, and defence lead the breakout categories. Deeptech now represents 36% of European VC (up from 11% in 2014). Defence tech funding up 148% YoY to $1.6bn, with AI x Defence leading at $929m (Helsing €600m, Harmattan $200m, Quantum Systems €340m). TechBio is maturing rapidly: 31 AI-discovered assets now in clinical phases, 30+ pharma contracts signed with TechBio companies in 2025, and 30% of all drugs discovered through R&D now involve AI. The Recursion-Exscientia merger signals the era of vertically integrated TechBio is beginning.

1.3. French Tech: the seed choke point

France decoupled from European recovery. Seed is the structural crisis: fewer seeds today means fewer fundable companies in 24 months.

€6.7bn raised (down 5% YoY), the only major ecosystem to decline while UK, Germany, and Nordics recovered. Paris captured 82% of funding, 67% of rounds.

Seed collapsed 38% (€0.8bn to €0.5bn) while growth rounds hit record highs (€3.5bn). The ecosystem is top-heavy: mega-rounds mask a hollowing out at the foundation.

This is a 2027 Series A crisis in the making. Structural risk for Series A/B pipeline in 2026 to 2027. Median time to Series A has increased 40%: from 1.5 years (pre-2021) to 2.1 years (2025).

Exits at five-year low: €5.3bn (down 65% YoY). Secondaries becoming the dominant liquidity path. Founders buying back companies (Fruitz, The Bradery). Unicorns becoming serial acquirers (Brevo, Doctolib, Contentsquare).

Talent is present; category leaders are scarce, for now. Numerous AI talents at Mistral, HuggingFace, Photoroom, Dataiku, etc. A few category leaders like Harmattan. But no skyrocketing apps yet (like Lovable, Legora, or n8n) or leading foundation models (like ElevenLabs). Early AI founders are still massively attracted to the US; YC is forcing Delaware incorporation. The next 18 months will be decisive.

For more details on French Tech, see Alexandre Dewez’s newly published annual report.

1.4. Technical undercurrents: signals from our community

What we are hearing from 300+ AI engineers and researchers across Reg.exe:

MCP reaching critical mass but fragmentation emerging: configs don’t transfer between tools. The standard is forming, but implementation is uneven.

Context engineering matters more than context windows. Teams now treat prompt files as critical infrastructure. The skill is shifting from “prompting” to “context architecture”.

Trust inflection in coding. Senior devs no longer read diffs from Sonnet 4.5. This is new. Code review is becoming exception-based, not comprehensive.

Evaluation remains unsolved. LLM-as-judge + human review feels inadequate for subjective tasks. Everyone knows this is the bottleneck; no one has cracked it.

Junior talent squeeze. Hiring is shifting from coding to architecture and prompting. The entry-level job market is compressing faster than expected.

Privacy trade-off still open. No one connects agents to Drive/Slack yet; waiting for the killer app that makes it worth the risk. Self-hosting and fine-tuning are back, driven by cost pressure, data sensitivity, and the pursuit of control.

Team.exe