The Investor

The Structural Shift

The venture capital stage taxonomy broke this week. When Thomas Dohmke (former GitHub CEO) raised a $60M seed at $300M valuation for Entire, it was the smallest of a cluster: Ineffable Intelligence (David Silver, AlphaGo architect) is targeting $1B at a $4B valuation — reportedly Europe's largest-ever seed. World Labs (Fei-Fei Li) closed $1B with AMD, NVIDIA, and Fidelity co-investing alongside a $200M Autodesk strategic. Mira Murati's Thinking Machines raised $2B. Saudi Arabia's Humain pumped $3B into xAI with a bizarre SpaceX share conversion mechanism. That's over $7B in announced AI capital in a single cycle.

The market has started calling these "coconut rounds" — a term coined by Bloomberg's Ed Ludlow — because the old vocabulary can't contain what's happening. The seed label now spans a 2,000x range ($1M to $2B), and the sole underwriting criterion for the mega end is founder pedigree.

Where Capital Is Fragmenting

Critically, this capital is not all flowing to the same thesis. The market is diversifying away from monolithic LLM scaling into distinct sub-categories, each attracting billion-dollar conviction bets:

Two patterns demand attention. First, the valuation floor for coconut rounds is ~$300M and the ceiling exceeds $4B — territory that used to require $50M+ ARR. Second, Silver's explicit framing — RL, not incremental LLM updates — is a thesis statement that the next capability discontinuity comes from a different paradigm than what 90% of AI capital currently funds. When competing chip companies (AMD and NVIDIA) both back World Labs, they're signaling spatial AI compute demand will be large enough to grow the pie for everyone.

Dohmke had the option to raise at $700M but deliberately settled at $300M — optimizing for dilution management and downstream optionality rather than maximizing Day 1 paper value. That discipline is rare and investable.

Portfolio Construction Implications

The elimination of staged financing removes the traditional risk-reduction mechanism that protected LPs. Companies attempting to compress 4-5 rounds over 7-10 years into 2-3 rounds over 3-5 years changes your reserve strategy, follow-on cadence, and expected hold period. Benchmark's 2020 fund sitting at 10x+ invested capital (driven by Legora, Mercor, Sierra, Cursor) and Jack Altman winding down a $275M fund to join as GP signals the top of the market believes AI-first bets are the generational trade.

The alpha is entirely in pre-announcement access. By the time a coconut round is reported, allocation is oversubscribed. The next 20 most likely coconut founders are senior leaders at OpenAI, Anthropic, Google DeepMind, Meta FAIR, and xAI who haven't yet left. The "disciplined coconut" tier ($50-100M seeds at $250-500M valuations) offers better risk-adjusted returns than $1-2B mega-seeds, where the path to a venture-returning outcome requires $20B+ exits.

The Physics You Can't Ignore

While billions pour into AI mega-rounds, the unit economics of actually running AI are governed by a physics constraint the market hasn't fully priced. During decode (token generation), the H100 runs at roughly 1% of its theoretical peak compute. Not because of bad software — because of memory bandwidth. Decode arithmetic intensity collapses to 1-4 FLOPs/byte, far below the H100's 295 FLOPs/byte threshold. And here's what makes this structurally investable: compute grows 3x every two years while bandwidth grows at half that rate. The problem gets worse with every GPU generation.

The KV cache is the binding constraint on AI unit economics. At 4K context, a 7B model serves 278 concurrent users per H100. At 128K context, that drops to ~8 users — a 35x cost escalation that most financial models ignore. Agentic systems, where agents share traces and context compounds, push every deployment toward the expensive end of this curve.

The 75-330x gap between compute floor and API pricing isn't margin — it's operational overhead. This gap is the addressable market for inference optimization companies.

Three Investable Wedges

1. Inference-Specific Silicon

Groq (massive on-chip SRAM), Cerebras (wafer-scale silicon), Etched (transformer-specific ASICs), and AMD's MI300A (92 FLOPs/byte vs. H100's 295) are attacking the structural memory-bandwidth bottleneck. The irony: Qualcomm's Snapdragon at ~1 FLOP/byte is actually better matched to decode's 1-4 FLOPs/byte arithmetic intensity than NVIDIA's H100. The compute gap is 20,000x, but the bandwidth gap is only 66x.

2. Edge AI Infrastructure

At 100M MAU with 500 requests/user/month, cloud API costs $11.25M/month versus $1.0M on-device — an 11x premium. On-device cost is flat regardless of usage because users already own the hardware. This means always-on AI features are economically impossible on cloud metering but viable on-device. Three independent developments are converging to make this real: YOLO26 eliminates post-processing complexity for edge object detection, SLMs use quantization for local inference, and Python 3.14's GIL removal unlocks true parallelism for edge workloads.

3. Utilization Optimization

Moving from 10% to 30% GPU utilization delivers a 3x cost improvement — more impactful than most architectural innovations. LangChain's coding agent jumped from Top 30 to Top 5 on Terminal Bench 2.0 with nothing more than a harness change — no model improvement. This proves the value layer is shifting from model providers to orchestration and harness engineering. Meanwhile, FloTorch's 2026 benchmark showed simple RAG chunking beating complex methods at 3-5x lower infrastructure cost, threatening the valuation thesis for over-engineered RAG tooling startups.

The memory wall, not the compute ceiling, determines who wins AI infrastructure; every dollar invested in faster FLOPs without proportional bandwidth is a dollar wasted on decode workloads.

The Convergence

Five independent signals this cycle point to the same conclusion: the AI application layer is being systematically commoditized by platform incumbents, and startups without distribution moats, proprietary data, or deep workflow integration are entering the kill zone.

Signal 1: Google's Lyria 3. Text-to-music, photo-to-music, and video-to-music generation integrated directly into Gemini with SynthID watermarking. Standalone AI music platforms like Suno and Udio now face a distribution problem they cannot solve — Google ships this to hundreds of millions of users at zero marginal cost.

Signal 2: GitHub Agentic Workflows. Microsoft is embedding agentic AI directly into the developer infrastructure layer where 100M+ developers already live. Developers describe desired outcomes in plain Markdown; agents handle triage, documentation, code quality, and more. This is textbook platform bundling that compresses the TAM for every standalone AI coding tool.

Signal 3: Disposable Interfaces. AI coding tools now enable end-users to generate custom interfaces that bypass traditional SaaS UIs entirely, accessing product value directly through APIs. If your portfolio company's moat is "beautiful UI" rather than proprietary data, that moat is dissolving.

Signal 4: Finance Copilot Graveyard. With 100+ LLM-enabled entrants fragmenting ARR, questionable unit economics from API dependence on Bloomberg/FactSet data, and LLM reliability gaps that institutional workflows cannot tolerate, the standalone finance copilot is a value trap. The real unlock — seamless two-way Excel integration — hasn't been cracked yet.

Signal 5: AI Monetization Fork. Perplexity pulled sponsored answers entirely, betting that trust drives subscriptions. OpenAI and Google are testing ads in AI responses. This divergence determines where marketing dollars flow — and whether AI ad-tech has a structural ceiling.

When Google gives away AI music generation for free inside a product with hundreds of millions of MAUs, quality parity is no longer a moat. The same logic applies to every AI feature that can be bundled into an existing platform.

Where Value Migrates

The value is migrating to three places: platform owners (Google, OpenAI, Microsoft), infrastructure providers (content provenance, agentic commerce rails, AI agent authorization), and frontier research labs with breakthrough capabilities platforms can't easily replicate. AI agent authorization is emerging as a particularly compelling new category — ReBAC systems like SpiceDB are structurally superior to static policy engines for agentic workloads, and the TAM expands with every enterprise AI agent deployment. OpenAI's hire of Charles Porch (15-year Meta/Instagram veteran) as VP of Global Creative Partnerships signals they're building cultural legitimacy as a moat — not just technology.

Meta's Bellwether Trial: The Most Underpriced Risk in Social Media

Mark Zuckerberg took the stand in a bellwether addiction trial where internal emails directly contradict his testimony. In 2015, Zuckerberg emailed teams to boost time spent by 12% as a 2016 goal — then testified Meta doesn't "give teams goals on time spent." Internal estimates showed 4 million children under 13 on Instagram while the company claimed they weren't allowed. Adam Mosseri said on a 2020 podcast that "there's such a thing as being addicted to a social media platform" — then testified he disagrees addiction exists.

TikTok and Snap settled before trial, the strongest signal that the industry views litigation risk as material. This is a bellwether case shaping thousands of pending lawsuits. If you have any portfolio exposure to ad-supported social platforms, you need a litigation scenario in your model — specifically, model 15-25% engagement reduction mandates.

Fed: Higher for Longer Is Now the Base Case

February 18 Fed minutes revealed a divided committee with rate cuts months away. The 10-Year Treasury ticked to 4.079%. Any DCF model assuming H1 2026 rate relief is stale. The implication for growth equity is direct: discount rates stay elevated, compressing present values of future cash flows across your portfolio.

Supply Chain Single Points of Failure

Trump invoked the Defense Production Act for glyphosate and elemental phosphorus, revealing the US has exactly one domestic producer of both chemicals — with China as the only alternative. Elemental phosphorus feeds into semiconductors, batteries, and military applications. Simultaneously, Bayer faces a three-body problem: a rumored $7.25B Roundup settlement, a pending Supreme Court case, and the DPA order that validates glyphosate as strategically essential. The Supreme Court ruling is the binary trigger — if for Bayer, the settlement closes and the stock re-rates; if against, liability widens.

Cybersecurity: Detection Moats Eroding

ETH Zurich demonstrated 25 attacks across Bitwarden (12), LastPass (7), and Dashlane (6) — breaking zero-knowledge encryption guarantees across 60M users. Separately, ADWSDomainDump bypasses both CrowdStrike Falcon and Microsoft Defender using ADWS (port 9389) instead of LDAP. ShinyHunters executed 15 breaches in 7 weeks of 2026 via social engineering and SSO abuse. The pattern: identity is the primary attack surface, and detection-based defenses are hitting structural limits. This creates displacement opportunities in next-gen credential management, ITDR, and AI agent security (Nono's kernel-enforced sandbox for MCP/LLM workloads).

The market is telling you AI infrastructure is investable, AI applications are fragile, social media faces a litigation repricing nobody's modeling, and the Fed just took H1 rate cuts off the table — adjust your discount rates and your deal flow filters accordingly.