AI Inference Prices Crash 90%, SaaS Margins Squeezed

◆ DEEP DIVES

1. 01

   The 90% Inference Price Collapse — Your Model-Layer Thesis Has 90 Days to Adapt

   <h3>The Commodity Tsunami</h3><p>ByteDance's <strong>Seed 2.0 Pro</strong> delivers frontier-class AI at <strong>$0.47 per million input tokens</strong> — a 73% discount to OpenAI's GPT-5.2 ($1.75) and a 91% discount to Google's Gemini 3 Pro ($5.00). This isn't a marginal price cut; it's an order-of-magnitude compression in a single competitive cycle. Seed 2.0 matches or beats both Western frontier models across math, reasoning, and vision benchmarks while being priced like a mid-tier API.</p><table><thead><tr><th>Model</th><th>Provider</th><th>Price/M Input Tokens</th><th>Benchmark Position</th></tr></thead><tbody><tr><td><strong>Seed 2.0 Pro</strong></td><td>ByteDance</td><td>$0.47</td><td>Matches/beats GPT-5.2 & Gemini 3 Pro</td></tr><tr><td><strong>GPT-5.2</strong></td><td>OpenAI</td><td>$1.75</td><td>Frontier; first autonomous physics discovery</td></tr><tr><td><strong>Gemini 3 Pro</strong></td><td>Google</td><td>$5.00</td><td>Frontier; Deep Think reasoning</td></tr><tr><td><strong>Minimax M2.5</strong></td><td>Minimax</td><td>Cost-optimized</td><td>RL-at-scale for agentic/coding</td></tr></tbody></table><h4>But Value Is Bifurcating, Not Disappearing</h4><p>While commodity inference races to zero, <strong>differentiated AI capabilities are becoming priceless</strong>. GPT-5.2 autonomously discovered and formally proved an original result in theoretical physics in 12 hours — verified by Harvard, Cambridge, and Princeton physicists. Harvard's Andrew Strominger said the AI "chose a path no human would have tried." This is AI's first original contribution to theoretical physics.</p><p>Simultaneously, the inference hardware war is creating its own investment layer. OpenAI deployed <strong>Cerebras chips for 15x speed gains</strong> (using a smaller model), while Anthropic achieved <strong>2.5x speed on full Opus 4.6</strong> without quality sacrifice. These divergent approaches — speed vs. quality — are segmenting the market and validating custom inference silicon as a standalone category.</p><blockquote>Commodity inference is racing to zero while differentiated AI capabilities are becoming priceless — the model layer is bifurcating, and your portfolio positioning must reflect this split.</blockquote><h4>The Microsoft-OpenAI Decoupling Signal</h4><p>Adding urgency: <strong>Microsoft is actively building in-house AI models</strong> under Mustafa Suleyman to reduce OpenAI dependency. This isn't a hedge — it's a strategic decoupling that threatens OpenAI's most valuable commercial relationship. OpenAI's deepening Azure data-layer lock-in (Cosmos DB for writes, co-developed replication features) creates a paradox: the infrastructure dependency deepens even as the commercial relationship frays.</p>

   Action items

   • Stress-test every portfolio company with foundation model API exposure against a 90% inference cost decline scenario by end of Q1
   • Reassess any direct or secondary OpenAI exposure by March 15, modeling Microsoft decoupling impact on revenue
   • Initiate diligence on AI-for-science vertical companies at Series A/B this quarter
   • Build an inference hardware thesis covering Cerebras, Groq, and Nvidia positioning by Q2

   Sources:🔬 GPT-5.2 makes an original physics discovery · OpenAI + OpenClaw 🤖, ChatGPT Lockdown Mode 🔒, inference speed tricks ⚡ · Compound engineering 🚀, OpenClaw founder joins OpenAI 💼, the AI vampire 🧛 · ByteDance Video AI 🎬, Designer Hiring Surge 📈, Airbnb AI Search 🏡

2. 02

   Agent Orchestration Is the Next Platform War — And the Security Gap Is the Greenfield Opportunity

   <h3>The Platform Layer Is Forming Now</h3><p>Seven independent sources converge on the same signal: <strong>AI agent orchestration is the contested next platform layer</strong>, and the competitive dynamics are already intense. OpenAI acqui-hired Peter Steinberger — creator of OpenClaw (120,000+ GitHub stars, 20,000 forks) — beating Meta in a competitive process. The deal's backstory is instructive: when OpenClaw first appeared as "ClawdBot," it <strong>defaulted to Anthropic's Claude</strong>, potentially driving millions of paying users to Anthropic. Anthropic's response? A cease-and-desist over the name. While Anthropic litigated, OpenAI negotiated.</p><table><thead><tr><th>Company</th><th>Agent Strategy</th><th>OpenClaw Outcome</th><th>Ecosystem Approach</th></tr></thead><tbody><tr><td><strong>OpenAI</strong></td><td>Acqui-hired Steinberger; OpenClaw stays open-source</td><td>Won</td><td>Ecosystem cultivation — Android model for agents</td></tr><tr><td><strong>Meta</strong></td><td>Competed to acquire/recruit</td><td>Lost</td><td>Aggressive but outbid</td></tr><tr><td><strong>Anthropic</strong></td><td>Sent cease-and-desist; lost distribution channel</td><td>Self-inflicted loss</td><td>Legal-first; destroyed ecosystem opportunity</td></tr></tbody></table><p>This mirrors the <strong>cloud middleware pattern from 2012-2016</strong>: compute commoditized, value accrued to orchestration (Kubernetes, Terraform, Docker). The agent orchestration tools emerging now — klaw for managing hundreds of agents, Warp's Oz platform, OpenClaw's foundation — are the Kubernetes of the agent era.</p><h4>The Security Gap Is the Bigger Opportunity</h4><p>Here's what makes this urgent: <strong>agent security infrastructure doesn't exist yet</strong>. 1Password's new SCAM benchmark tested all 8 frontier AI models and every single one failed critical security tests — entering credentials on phishing pages, forwarding passwords to external parties. Safety scores ranged from <strong>35% to 92%</strong> with critical failures across the board. OpenClaw operates with the same permissions as the user who installs it. ByteDance demonstrated 96-step autonomous CAD workflows. ElevenLabs is enabling outbound AI sales calls.</p><p>The attack surface is massive and growing. Meanwhile, 64.4% of product roadmaps include agentic AI, 67% of teams claim to be building agentic workflows, and 85% expect it to be <strong>table stakes within three years</strong>. But there's a critical caveat: teams have "wildly different definitions of 'agentic'" — the classic pre-disillusionment signal that predicts a 12-18 month shakeout.</p><blockquote>Agent orchestration is the cloud middleware opportunity of the AI era — but the security gap means enterprises cannot deploy agentic AI safely, creating a greenfield category analogous to cloud security in 2014.</blockquote>

   Action items

   • Source 3-5 Series A/B agent security startups building sandboxing, credential scoping, and behavioral monitoring by end of Q1
   • Map the agent orchestration competitive landscape (klaw, OpenClaw foundation, Warp/Oz) and identify seed-stage entry points this quarter
   • Audit portfolio DevOps companies for credible agentic AI strategy — escalate to board level for any company without one
   • Discount self-reported 'agentic AI' claims in diligence — require demos of autonomous decision-making within constraints

   Sources:🔬 GPT-5.2 makes an original physics discovery · OpenAI + OpenClaw 🤖, ChatGPT Lockdown Mode 🔒, inference speed tricks ⚡ · Community Trust Management 🎫, Java's Debt Wall 🧱, AI Tool Surge 📈 · 300 Chrome Extensions Caught Stealing 🥷, Product Engineering & Supply Chain 🚚, Snail Mail Attack on Crypto Users ✉ · AI teams, adoption, and public reading, · ⚕️ MODESTLY ☘ Monday, February 16, 2026 ☘ C&C NEWS 🦠

3. 03

   Per-Seat SaaS Is Structurally Breaking — The Botkeeper Funeral, the Ramp Playbook, and the Stripe Validation

   <h3>Three Data Points, One Thesis</h3><p>The per-seat SaaS model that generated the last decade's venture returns is entering structural decline, and three events this week draw the line with unusual clarity.</p><p><strong>First: Botkeeper shut down after $90M raised and 11 years.</strong> Despite building AI that coded 80%+ of accounting transactions accurately, it never achieved the operational embedding that creates switching costs. On the same day, <strong>Ramp launched an Accounting Agent</strong> claiming 3.5x more auto-coded transactions and 98% sync accuracy — bundled with its existing expense management platform. The lesson is brutal: AI capability is table stakes; the moat is owning the data loop and system-of-record position.</p><p><strong>Second: Stripe paid $1 billion for Metronome</strong> because its core billing architecture fundamentally cannot handle AI-era usage-based pricing. Stripe's existing system relies on pre-aggregated data pushed via HTTP — unsuitable for event streaming and real-time metering. Rebuilding internally would have been a multi-year breaking change. This is a build-vs-buy capitulation that <strong>validates the entire usage-based billing category</strong> and sets the valuation floor for remaining independent players (Orb, Amberflo, m3ter).</p><p><strong>Third: $470B+ in hyperscaler AI infrastructure spend</strong> is being pulled directly from enterprise software budgets — not incremental IT spend. If 10 AI agents do the work of 100 sales reps, you don't need 100 Salesforce seats. Spotify's CEO confirmed top developers haven't written a single line of code in 2026. Goldman Sachs embedded Anthropic engineers for 6 months on trade accounting. The procurement model is shifting from vendor contracts to co-development partnerships.</p><table><thead><tr><th>Signal</th><th>What Died</th><th>What Won</th><th>Moat Type</th></tr></thead><tbody><tr><td>Botkeeper vs. Ramp</td><td>AI wrapper ($90M, shutdown)</td><td>Embedded system of record</td><td>Data gravity + distribution</td></tr><tr><td>Stripe/Metronome</td><td>Subscription billing architecture</td><td>Usage-based metering ($1B exit)</td><td>Infrastructure lock-in</td></tr><tr><td>Goldman-Anthropic</td><td>SaaS vendor contracts</td><td>6-month co-development partnerships</td><td>Operational embedding</td></tr></tbody></table><blockquote>In vertical AI, the dispatcher is dead and the embedded system of record is the only investable position — Botkeeper's $90M funeral and Ramp's launch on the same day is the market drawing the line for you.</blockquote>

   Action items

   • Audit every portfolio company's revenue exposure to per-seat pricing and initiate pricing model migration conversations with management teams by end of Q1
   • Map the usage-based billing infrastructure landscape and identify remaining independent players as M&A targets by March 30
   • Apply the Botkeeper 'dispatcher problem' framework to every vertical AI company in pipeline — test whether they own the data loop or are simply dispatching LLM calls

   Sources:AI acqui-hire wave 🤝, secondary markets boom 📊, token anxiety 🧠 · Coinbase surges 📈, Goldman Sachs uses Claude 🤖, Ramp's Accounting Agent 👨‍💼 · Compound engineering 🚀, OpenClaw founder joins OpenAI 💼, the AI vampire 🧛 · ChatGPT's first ads 🛒, 7 growth mistakes 👎🏼, Claude's download surge 🔼

4. 04

   The Private-Public Valuation Gap Is Widening — Late-Stage Discipline Required Now

   <h3>The Numbers Tell a Dangerous Story</h3><p>AI startups are reaching <strong>$1B valuations in 3.4 years</strong> — half the 7-year historical average. Secondary market volume hit <strong>$3.5B in 2025</strong>, up 75% YoY. Cursor, Perplexity, and ElevenLabs all reached liquid secondary markets within 3 years of founding. The private market is sprinting.</p><p>The public market is saying "not so fast." <strong>70% of 2025's largest IPOs are trading below their offering price</strong> — 14 of the top 20. Figma, the poster child of the 2025 IPO class, is down <strong>80%+ since its July IPO</strong>. Clear Street slashed its IPO target from $1.05B at $11.8B valuation to <strong>$364M at ~$7.2B</strong> — a 39% valuation haircut and 65% reduction in raise size.</p><table><thead><tr><th>Metric</th><th>Current</th><th>Historical</th><th>Delta</th></tr></thead><tbody><tr><td>Time to $1B valuation</td><td>3.4 years</td><td>7 years</td><td><strong>-51%</strong></td></tr><tr><td>Secondary market volume</td><td>$3.5B (2025)</td><td>$2.0B (2024)</td><td><strong>+75%</strong></td></tr><tr><td>Top IPOs below offering</td><td>14 of 20</td><td>—</td><td><strong>70% failure rate</strong></td></tr><tr><td>Figma post-IPO</td><td>$22.53</td><td>IPO price ~$110+</td><td><strong>-80%+</strong></td></tr><tr><td>Clear Street IPO haircut</td><td>$7.2B</td><td>$11.8B prior round</td><td><strong>-39%</strong></td></tr></tbody></table><h4>The Macro Backdrop Compounds the Risk</h4><p>A growth-to-value rotation is accelerating: <strong>Dow +2.99% vs. Nasdaq -2.99% YTD</strong>, with Bitcoin down 21.91%. The 10-Year Treasury at 4.056% keeps pressure on duration-sensitive growth names. Coinbase rallied 20% on a 20% revenue decline — the market is pricing regulatory optionality (stablecoin legislation, Fed payment rail access), not fundamentals. This is a market that rewards tangible catalysts and punishes growth-without-profitability.</p><h4>The Stealth M&A Complication</h4><p>Adding complexity: <strong>79% of 5,700 AI acquisitions between 2020-2025 were undisclosed</strong> — 4,500 quiet talent grabs invisible to standard deal databases. The 75th percentile AI deal size <strong>tripled from $82M to $248M</strong>. Anyone relying on PitchBook or Crunchbase for AI sector comps is working with 21% of the data. This information asymmetry distorts both entry valuations and exit expectations.</p><blockquote>Private markets are pricing AI companies for hypergrowth while public markets reject those valuations at the IPO window — the 70% IPO underperformance rate isn't a blip, it's the exit environment telling you the markup you're underwriting may not exist.</blockquote>

   Action items

   • Cap secondary market entry at 20x forward ARR and require demonstrated path to public-market-viable unit economics for all new positions
   • Reassess late-stage fintech markings using Clear Street's 39% haircut as a new comparable by next IC meeting
   • Build a proprietary AI acqui-hire tracker using alternative data sources to capture the 79% of undisclosed transactions
   • Monitor Figma earnings Wednesday for read-through on SaaS growth multiples and IPO class health

   Sources:AI acqui-hire wave 🤝, secondary markets boom 📊, token anxiety 🧠 · Coinbase surges 📈, Goldman Sachs uses Claude 🤖, Ramp's Accounting Agent 👨‍💼 · ⚡ Crisis of memory · Ethereum Leadership Change 🏛️, Everything is Market 💹, Solana 2026 🗓️