ClarityBeta

Product & Strategy

The Product Desk

6 min read · 1,151 words · 3 moves to watch
The Signal

Jack Dorsey told JPMorgan's elite Tech100 that using AI coding agent Goose every morning

When C-suite executives personally adopt coding agents and start doing headcount math, reorgs follow within quarters, not years. If you aren't proactively modeling your team's AI-augmented productivity for leadership right now, someone above you will do it with cruder math and less nuance.

In Play

  1. CEO Coding Agent Adoption → Workforce Restructuring

    Dorsey (Block) and Ghodsi (Databricks) both publicly stated AI coding agents prompted them to rethink headcount. HashiCorp's Hashimoto runs agents in parallel all day. The pattern: when CEOs prototype with agents, reorg conversations follow within quarters.

    Ask Clarity

  2. AI Spend Credibility Wall — Markets Demanding ROI Proof

    Microsoft's worst quarter since 2008 (down 34%) and Nasdaq in correction (down 11%) signal the 'invest now, monetize later' narrative hit its wall. Yet SoftBank secured $40B for OpenAI and Moonshot raises at $18B. Public markets punish; private capital floods in. Your AI budget pitch needs hard unit economics now.

    Ask Clarity

  3. Anthropic's Compounding Vendor Risk — Capybara, Throttling, and 250M New Users

    Anthropic leaked Capybara (a tier above Opus with substantially better coding/reasoning scores), is throttling existing API customers, AND just licensed tech to Yahoo Scout's 250M users. Cybersecurity stocks slid on the Capybara rumor alone. You're now competing for Anthropic compute with Yahoo's entire user base.

    Ask Clarity

  4. H100 Prices Defy Depreciation — Compute Cost Models Are Broken

    H100 rental prices reversed their depreciation curve since December 2025 and are now worth more than at 2022 launch. Chip shortage plus agent/reasoning demand inflection killed the declining-cost assumption. Any cost model built on 2024 GPU pricing is materially wrong.

    Ask Clarity

  5. RSA 2026: Security Products Collapse Into API Calls Within 1-3 Years

    Daniel Miessler spoke with ~50 RSA vendors and found most building proprietary AI dashboards — the wrong play. His thesis: all security products become API calls consumed by customer-controlled agentic orchestration within 1-3 years. The question for any B2B PM: can your product's core value be consumed headlessly?

    Ask Clarity

Deep Dives

The Dorsey Moment: CEOs Are Doing AI Headcount Math — And You Need to Do It First

What Happened at Tech100

At JPMorgan's invitation-only Tech100 conference (March 25-27, Yellowstone Club, Montana), Jack Dorsey told investors that using AI coding agent Goose every morning led him to conclude he could nearly halve Block's workforce. Databricks CEO Ali Ghodsi described the identical realization hitting his team. These aren't startup founders hypothesizing — they're CEOs of major companies telling their largest investors, on the record, that AI agents have changed their headcount assumptions.

When C-suite executives personally adopt AI coding tools and start doing mental math on headcount, reorg conversations follow within quarters, not years.

The Cross-Source Pattern

This isn't isolated executive enthusiasm. HashiCorp co-founder Mitchell Hashimoto describes running AI agents constantly in the background as his production workflow — when he codes, they plan; when they code, he reviews. The emerging agent UX pattern across coding tools is converging on 'fleet management for software' — kanban-like cards, isolated worktrees, agent-owned tasks, and diff-based review. CursorBench data shows a median of 181 lines changed per task in real-world agent sessions. OpenAI's Codex is building a plugin ecosystem around this paradigm.

The Critical Contradiction

Here's the tension every PM must internalize: research shows AI tools increase competition entry by 42% without improving individual success rates. CEOs are seeing dramatic personal productivity gains and extrapolating to workforce-wide cuts, but the evidence suggests AI democratizes participation rather than multiplying quality. The PMs who thrive in this environment are those who can articulate the nuanced reality: AI agents change what your team works on, not just how many people you need.

Why You Must Move First

The AI job market already shows 3.2 open roles per qualified candidate, with most applicants lacking critical skills. You can't hire your way to faster delivery. But if you wait for your CEO to have their own 'Goose morning' and dictate cuts from the top, you lose the ability to shape the outcome. The PM who proactively models team productivity with agents — and proposes a reallocation plan that protects your strongest engineers while capturing the gains — controls the narrative.

What to do

  1. Instrument your team's AI coding tool usage and quantify the productivity delta (tasks completed, time-to-merge, lines per session) over the next two sprints

  2. Evaluate Goose (Block's coding agent) alongside your current AI dev tools stack this sprint — if Dorsey's drawing workforce conclusions from it, you need direct experience

  3. Draft a 'team rebalancing' proposal by end of Q2 that shifts engineer hours from code production to specification, review, and agent orchestration

The AI ROI Credibility Wall — Public Markets Are Done Waiting

The Numbers Tell a Brutal Story

Microsoft's stock is down 34% since October — its worst quarter since 2008. The Nasdaq 100 is in correction territory, down 11% from its peak, with Business Insider attributing part of the decline to 'AI anxiety.' The S&P 500 just posted five consecutive down weeks, the longest streak since 2022. Investors are described as 'recoiling' from continued AI infrastructure spending without clear returns — and a new fear is driving the selloff: that AI startups building agents will replace, not augment, incumbent software products like Microsoft's.

The Capital Contradiction

Here's what makes this moment unprecedented for product leaders: public markets are punishing AI spend while private capital floods in. SoftBank just secured a $40B unsecured bridge loan from JPMorgan and Goldman Sachs to fund its OpenAI commitment. Moonshot AI is raising at $18B. The implication for your planning: your company's budget may tighten (correction-driven caution), but your AI-focused startup competitors will remain exceptionally well-funded.

The market is no longer differentiating between 'doing AI' and 'getting ROI from AI.' Your next AI feature pitch needs unit economics, not a demo.

The Macro Compressor

Rate expectations have undergone the most dramatic reversal in recent memory: CME FedWatch data shows markets went from pricing 90% probability of rate cuts by September to 52% probability of a rate hike this year — in a single month. Oil at $110/barrel from the Iran conflict is the primary driver. For product leaders, this directly impacts enterprise deal velocity, customer willingness to spend on new tools, and internal headcount budgets by Q3.

What 'AI Productivity' Actually Means

The research finding that AI tools increase competition entry by 42% without improving individual success rates is the data point every PM needs when framing AI feature value. If you're telling leadership 'AI will improve user outcomes,' the evidence says otherwise. AI lowers barriers to participation. The winning products measure adoption, activation, and time-to-first-value — not quality improvement claims that don't hold up to scrutiny.

What to do

  1. Rewrite your top AI feature's business case with hard ROI metrics — cost savings per user, revenue per AI-assisted conversion, or churn reduction with confidence intervals — before your next planning review

  2. Accelerate any enterprise deals planned for Q3 into Q2 wherever contract flexibility allows

  3. Reframe AI feature OKRs from 'improved outcomes' to 'expanded access and engagement' based on the 42%/0% research, and update PRD metrics sections accordingly

Anthropic's Triple Risk Event — Capybara, Throttling, and 250M New Mouths to Feed

The Capybara Leak Changes the Frontier Calculus

Anthropic accidentally exposed its most powerful unreleased model through an unsecured data cache. Capybara sits above Opus as a new tier — 'larger and more intelligent' than Claude Opus 4.6, with substantially better coding, academic reasoning, and cybersecurity benchmark scores. Fortune corroborated the leak. The cybersecurity capabilities alone were enough to crater security stocks on March 27, signaling that markets now price in AI labs commoditizing entire vertical software categories with a single model release.

The Capacity Squeeze Is Real

Anthropic is simultaneously throttling existing Claude API customers while licensing technology to Yahoo Scout's 250 million users. If you're building on Anthropic's API, you're now competing for compute with Yahoo's entire installed base plus Claude's organic consumer demand. Widespread 529 errors were reported during the leak period. Google is reportedly close to funding Anthropic's data center buildout, which eventually solves the capacity problem — but not on your Q2 timeline.

Capybara will exist as a capability but be scarce and expensive. Products that can integrate quickly and absorb the premium get a temporary moat. Products that can't need a fallback plan now.

The Open Model Escape Valve

The good news for build-vs-buy: the open-closed gap has narrowed dramatically. Zhipu's GLM-5.1 scored 45.3 on coding benchmarks vs. Claude Opus 4.6's 47.9 — a 5.4% gap, down from 26% with the prior GLM-5 (35.4). Quantization breakthroughs like TurboQuant now enable running Qwen 3.5-9B on a standard MacBook Air (M4, 16GB) with 20K context. RotorQuant achieves 10-19x speed improvements at 0.990 vs. 0.991 cosine similarity — essentially equivalent quality. The Arena leaderboard confirms the open-closed gap is 'much narrower than a year ago.'

Your Vendor Strategy Decision Matrix

ScenarioRiskAction
Anthropic throttling worsensFeature degradation, SLA breachMulti-model fallback to OpenAI/Google
Capybara launches, capacity-constrainedCompetitors with access get temporary moatEarly access request + premium budget allocation
Post-IPO pricing increaseUnit economics compressionOpen model evaluation for non-frontier tasks
Government standoff restricts capabilitiesFeature removal or access restrictionAbstraction layer enabling rapid model swap

What to do

  1. Implement multi-model fallback architecture this sprint — map which features break during Anthropic throttling and configure automated failover to OpenAI or Google endpoints

  2. Run a structured evaluation of open models (Qwen 3.5-35B, GLM-5.1) against your current Anthropic usage for your top 3 use cases within 30 days

  3. Run a 'commoditization stress test' on your roadmap: for every major feature, ask what happens if a foundation model ships this as a native capability in 6 months

The bottom line

Tech CEOs are personally using AI coding agents, doing headcount math, and concluding they can halve their workforces — while public markets just posted the worst tech quarter since 2008 demanding AI ROI proof, and Anthropic is simultaneously leaking a step-change model, throttling existing customers, and licensing to 250 million Yahoo users. The PMs who survive this compression aren't building more AI features; they're proactively modeling AI-augmented team productivity before leadership does it for them, hardening vendor diversification against a capacity-constrained Anthropic, and replacing every 'AI will improve outcomes' claim with unit economics — because the evidence says AI expands participation 42% without improving individual results.