MY TAKE
Three threads crossed this week that, seen together, draw the strategic map of what’s coming.
The first thread is institutionalized AI asymmetry. Anthropic has made public that Claude Mythos can autonomously discover zero-days, has reserved it for 40 organizations inside Project Glasswing, and Dario Amodei was seen meeting with the White House chief of staff and the Treasury secretary.
In parallel, power users and developers were reporting that Claude in production was degrading. I noticed it a lot myself, but I figured it was the usual thing that happens whenever they ship a new model, where they seem to divert capacity.
No. Opus 4.7 is worse.
The IMPORTANT point here is that the best model is no longer sold: it’s allocated. And when a model capable of finding zero-days is allocated rather than sold, we’re talking about a state-grade asset, not a SaaS product.
I have to tell you, I saw this coming and frankly they were overdue… Did you really think superintelligence was going to be for everyone?
Ha!
The second thread is that the old-guard founders are stepping away from the helm. Reed Hastings has announced he’s leaving the Netflix board after 29 years, coinciding with the Q1 results.
Armon Dadgar, HashiCorp’s last technical co-founder, left this week, less than two years after the IBM acquisition closed. Jack Dorsey published in Fortune the details of his thought process behind cutting 40% of Block.
Every case has its context, I get it… but the pattern repeats:
the era of the anchor founder is closing across several flagship companies. Capital allocation decisions no longer pass through their filter. This has direct consequences for how boards will approve AI investments in the coming quarters: without a founder, the CFO and the activists will gain weight.
The third thread is the numbers… TSMC reported Q1 with profit +58% YoY, 66.2% gross margin, and raised its 2026 outlook to more than 30% growth. HPC (AI + 5G) is already 61% of its revenue.
Netflix beat estimates and the stock dropped between 8% and 10%… Does that behavior ring a bell from another one of my newsletters?
Goldman booked $1.5 billion in advisory fees (+89% YoY), a signal that the M&A wave is not noise. These three data points say the same thing: capital is concentrating in infrastructure and in consolidation transactions, not in second-tier organic growth.
I think the week of April 29, with Alphabet, Meta, Microsoft and Amazon all reporting on the same day, is going to be the test. If AI capex keeps growing while cloud revenue decelerates, the market conversation is going to change fast. And the founders will no longer be there to hold up the narrative…
What a ride we’re in for… if this is enough for you, I’d ask you to comment or share, and if you want more detail I’ll leave you with my minions.
THE BOMBSHELL OF THE WEEK
Anthropic Mythos: the model the government doesn’t want you to see
The most important story of the week has layers the mainstream press is processing only superficially. Let’s reconstruct the sequence:
April 13. Fortune reports that Anthropic has detected that Claude Mythos can autonomously discover thousands of zero-day vulnerabilities in major operating systems and browsers — flaws human researchers hadn’t found after years of work. The news triggers panic across the cybersecurity industry.
April 14. Anthropic simultaneously faces a second crisis: high-volume users and developers report en masse that Claude’s performance has fallen off a cliff. Threads on Hacker News, Reddit and X describe a model that “doesn’t follow instructions,” takes “inappropriate shortcuts” and fails on complex workflows. The dominant hypothesis among users: Anthropic has redirected compute toward Mythos, deliberately degrading the models in production.
April 16. Anthropic launches Claude Opus 4.7 with general availability, at pricing identical to 4.6 ($5 / $25 per million tokens). The company publicly admits that Opus 4.7 does not match Mythos’s capabilities. A new record in corporate transparency: telling the market you have a superior product you’re not going to sell. The same day, Axios ran the headline: “Anthropic’s AI downgrade stings power users”.
April 17. CNBC reports that Dario Amodei met with Susie Wiles (White House chief of staff) and Scott Bessent (Treasury secretary) to discuss Mythos. Trump tells reporters he “had no idea” Amodei was there.
Anthropic has created the first institutionalized two-speed AI model: a product for the public (Opus 4.7) and a strategic weapon for 40 selected organizations inside Project Glasswing (Microsoft, Apple, Google, CrowdStrike, JPMorgan and 35 others). This is not a soft launch. It’s a deliberate bifurcation between commercial AI and state-grade AI.
That Bessent (Treasury) is in the room and not Rubio (State) or the DNI suggests the White House is treating this as a financial-strategic asset, not a traditional national security threat.
The signal: if your company isn’t in the Glasswing 50, you’ve just been left out of the first tier of cyber defense for the coming era. The access asymmetry between companies will be brutal, and the conversation many boards will have in Q2 is whether they need to lobby to get on the list or build partnerships that give them indirect access.
POWER MOVES
Reed Hastings leaves the Netflix board
On April 16, coinciding with the publication of Q1 2026 results, Netflix announced that Reed Hastings, co-founder and chairman, will leave the board at the annual meeting on June 4. 29 years after founding the company, Hastings exits with no founder-surname successor. Co-CEOs Ted Sarandos and Greg Peters are left in charge without founder oversight for the first time in the company’s history.
There was no prior public drama. Hastings has spent two years scaling back his active role, sold his position in Powder Mountain (Utah) and is focused on philanthropy. The stock fell between 8% and 10% in after-hours even though the quarter’s numbers beat estimates. The market wasn’t punishing the results: it was punishing the combination of conservative guidance and the founder’s exit.
The buried strategic signal: Netflix has gone three years without publishing quarterly subscriber data. Hastings was the architect of that opacity decision. With him gone, there’s no guarantee that policy survives pressure from activist investors. If Netflix starts reporting subscribers again, the first post-Hastings number will be the market event — not his departure.
Armon Dadgar exits HashiCorp
This week, Armon Dadgar, co-founder and CTO of HashiCorp (acquired by IBM for $6.4 billion in 2024), left the company. He’s moving from Seattle to New York. It’s the second founding departure: Mitchell Hashimoto had already left. HashiCorp as an entity inside IBM is left without either of its two original technical founders less than two years after the close.
The pattern matters more than the individual event. Technical co-founders in tech acquisitions last an average of 18–24 months post-close. IBM acquisitions historically have the worst track record for retaining founding talent. For any CTO considering selling their company to a large integrator, HashiCorp is the updated case study.
Jack Azagury takes the helm at Insight Enterprises
Effective April 13, Jack Azagury (29 years at Accenture, former Group Chief Executive of Consulting) takes over as CEO of Insight Enterprises, a Fortune 500 IT consulting firm with more than $14 billion in annual revenue. He replaces Joyce Mullen, who is retiring.
The hire points to an explicit “AI-first” strategy: the big IT integrators are accelerating their repositioning as AI transformation consultancies. The move reads best in parallel: Accenture has spent 18 months losing senior talent to smaller rivals and to PE funds consolidating integrators. Insight just landed one of the operators with the densest global pipeline in the sector.
MONEY TALKS
TSMC Q1: the bottleneck that dictates prices to Nvidia and Apple
Reported on April 16, TSMC’s quarter is the most relevant metric of the week for the entire AI sector:
- Revenue: $35.71 billion, +35% YoY, an all-time record
- Net profit: +58% YoY
- Gross margin: 66.2% (a historic extreme)
- HPC (AI + 5G): 61% of total revenue
- Advanced nodes (3nm + 5nm): 74% of wafer revenue
- Q2 guidance: $39.0–40.2 billion, +10% sequential
- 2026 outlook revised upward to more than 30% annual growth in dollars
The read nobody says explicitly: with these margins, TSMC has more pricing power over Nvidia, Apple and AMD than any regulator. AI customers are price-inelastic. CEO C.C. Wei literally said AI demand is “extremely robust” and that recent geopolitical conflicts “failed to dent” it. In an environment of active tariffs, holding a 66.2% margin means TSMC is passing part of the tariff cost on to its customers with barely any negotiating resistance.
Goldman Sachs: M&A is truly back
Goldman reported on April 13 (Sunday) before markets opened: $17.23 billion in revenue (second-best quarter in its history), EPS of $17.55 versus an estimate of $16.49, and advisory revenue of $1.5 billion (+89% YoY). AUM at a record $3.7 trillion.
The 89% advisory figure is the one that matters for deal context: it confirms that the M&A machinery that has been reactivating for 18 months is not noise — it’s structure. Investment banks don’t hire 200 MDs without a real pipeline. For any CTO or founder, the direct implication: due diligence processes and transaction multiples are returning to 2021 levels, but with 2026 quality criteria (revenue quality, margin reality, AI defensibility).
Hexagon buys Waygate from Baker Hughes for $1.45 billion
Announced April 13. Waygate is the world leader in industrial non-destructive testing (turbines, aerospace, pipelines, critical assets): 1,500 employees, $630 million in annual revenue, headquartered in Germany. Close expected in H2 2026 pending regulatory review.
It’s not pure tech, but it’s the most relevant industrial acquisition of the week: Hexagon consolidates its position in industrial inspection with physical sensors, a layer that complements its measurement software. The useful data point for executives in heavy industry: $1–2 billion deals in industrial tech are getting done at revenue multiples between 2.2x and 2.5x, consistent with the pace Goldman confirms in its quarter.
Slate Auto: $650 million Series C, Bezos behind it, 160,000 reservations
On April 13, Slate Auto closed a $650 million Series C led by TWG Global (Mark Walter and Thomas Tull). Total raised: about $1.4 billion. An attack on the low end of the EV market with an electric pickup starting at $25,000. A $400 million investment in a plant in Warsaw, Indiana, with more than 2,000 jobs.
The implicit bet: Bezos against Tesla. With Musk focused on SpaceX, xAI and politics, Tesla is neglected as a product in the low-end segment. The timing relative to Trump’s tariffs (which punish Asian manufacturing) is strategically advantageous and explains why the funds accept this valuation.
Mintlify: $45 million Series B at a $500 million valuation
On April 14, Mintlify announced a $45 million Series B led by a16z and Salesforce Ventures, at a $500 million valuation. 20,000 corporate customers, 100 million annual documentation users.
The thesis: documentation as infrastructure for AI agents. If agents are going to consume APIs and codebases, they need structured, machine-readable documentation, not PDFs. That Anthropic uses Mintlify for Claude Code is the key endorsement. A $500 million valuation for a documentation company would have been unthinkable in 2022; in 2026 it’s plausible if you accept that the “knowledge layer” is critical for the agent era. The signal for VCs: the agent-native infrastructure layer is attracting multiples comparable to what data attracted in 2018–2020.
PRODUCT SECRETS
Claude Opus 4.7: radical transparency or admission of defeat
Anthropic launched Opus 4.7 on April 16 with general availability, pricing identical to 4.6, and a public note admitting it doesn’t match Mythos’s capabilities. For a CTO, the launch has two readings:
Reading 1 (deliberate transparency): Anthropic is building credibility against OpenAI and Google by differentiating on disclosure of limitations. It’s consistent with its “safety first” positioning.
Reading 2 (crisis containment): Opus 4.7 was pulled forward on the roadmap to respond to the developer backlash of the week of April 14. The pricing identical to 4.6 suggests the company doesn’t want to communicate an upgrade: it wants to communicate a repair.
Both readings can be true simultaneously. What’s true regardless of the motivation: the market now knows Anthropic has a superior model it doesn’t sell. That changes the negotiation of every enterprise contract: “why are you selling me Opus 4.7 if you have Mythos?” is the question every procurement team is going to ask in Q2.
Ghostwriter and the “forward-deployed engineer” problem
Sierra (Bret Taylor) launched Ghostwriter in early April and this week the discussion consolidated in enterprise AI channels after the first adoption data: Sierra is at $100 million ARR with a $10 billion valuation, but it requires “forward-deployed” engineers updating each enterprise customer’s agents in an almost artisanal way. The question analysts started asking this week: how many engineering FTEs per customer does it take to keep an agent in production?
The answer will determine whether enterprise AI has SaaS margins or hides a consulting model inside software pricing. Mintlify and Sierra share this problem: both are products that can scale or can turn into expensive consultancies in disguise. The market hasn’t decided yet.
REAL NUMBERS
Layoffs, week of April 11–18:
- Snap: 1,000 employees (16% of the workforce), 300 open positions closed, $500 million in projected annualized savings (April 15)
- Disney: approximately 1,000 employees across TV, film, marketing and corporate under new CEO Josh D’Amaro (April 14–15)
- Qualcomm: ~66 employees across 11 San Diego locations, in IT, engineering and cybersecurity divisions (WARN notice, California)
- Sama (Meta’s data-labeling contractor): more than 100 workers in Kenya
- Kintsugi: total shutdown confirmed this week; the Berkeley startup spent $16 million over 4 years of the FDA De Novo process without clearance
US tech layoffs — week ended April 15: 716 people confirmed (Computerworld/TrueUp).
2026 YTD as of April 17: 99,283 tech workers laid off. We’ll pass 100,000 before April is over.
Financial results reported this week:
Company — Q1 Revenue — YoY — Key figure: TSMC — $35.71B — +35% — Gross margin 66.2%. Goldman Sachs — $17.23B — — — Advisory +89% YoY. Netflix — $12.25B — +16% — Ad revenue on track for $3B annually.
The AI quality crisis nobody wants to admit (Lightrun 2026):
- 43% of AI-generated code changes require manual debugging in production after passing QA and staging
- PRs with AI-assisted code have 1.7x more issues than human code
- The incidents-to-PR ratio rose 242.7% as teams move from low to high AI adoption
- 0% of engineering leaders describe themselves as “very confident” that AI-generated code will behave correctly in production
Implication: the “code commits generated” ratio several CEOs are using as a productivity metric is capturing volume but not quality. Incidents-to-PR is the figure CFOs will start demanding in Q2.
THE DRAMA
Anthropic vs its own users: the silent “nerf”
The week of April 14, Anthropic faced a credibility crisis with its most valuable base: developers and power users paying for premium plans.
The narrative in technical communities: Claude has been deliberately “nerfed.” The model doesn’t follow complex instructions, takes shortcuts where it used to complete tasks. The users’ hypothesis is that Anthropic is redirecting compute toward Mythos and frontier projects, degrading the quality of the commercial product.
The official response was public but oblique. They neither confirmed nor denied the compute reduction. The launch of Opus 4.7 on April 16 at pricing identical to 4.6 was partly read as a reactive response. Axios headlined bluntly on April 16: “Anthropic’s AI downgrade stings power users”.
The real risk: Anthropic makes most of its money from enterprise and API. If developers perceive systematic degradation without transparent communication, the churn toward Gemini 3.1 Pro or models like Mistral will show up in the Q2 data. That kind of reputational damage takes quarters to repair.
Kintsugi: the health AI textbook case
The Berkeley startup (voice-based depression detection) shut down definitively this week. Seven years, $16 million spent on the FDA De Novo process over four years, no clearance and no additional funding. CEO Grace Chang made the decision to open-source all the code and pivot toward deepfake detection (same technology, zero regulation).
It’s the case study for why health AI needs a regulatory runway 2–3x that of other verticals. VCs with five-year horizons can’t fund seven-year FDA cycles. The implication for founders in regulated sectors: structure your funding and roadmap assuming the regulator is an execution variable, not a formality. For LPs with health AI exposure, Kintsugi forces a review of return-window assumptions.
THE WEEK AHEAD
April 22 — Tesla Q1 2026 earnings (after close). The most important report of the week. Context: Q1 deliveries were 358,023 units, -14% sequential, below estimates. The market wants answers to three questions: (1) when do robotaxis reach real revenue, (2) whether AI5 (the next onboard compute platform) is on schedule, (3) whether margins hold after the aggressive discounts Tesla used to move inventory in Q1. Musk has been doing Tesla events again in recent weeks, probably building the narrative for the earnings call.
April 22–24 — Google Cloud Next 2026, Las Vegas (Mandalay Bay). Google’s most important product-announcement event of the year. Declared focus: “AI agents in production at enterprise scale.” Expect news on Vertex AI, Gemini for Google Cloud and Ironwood infrastructure updates. The timing is strategic: Google Next happens one week before Alphabet’s earnings (April 29), which usually means the technical announcements support the narrative the company wants analysts to have in mind when hearing the results.
April 29 — Mega-Earnings Day. Four hyperscalers reporting on the same day:
- Meta Platforms: Q1 2026 after the close
- Amazon: Q1 2026
- Alphabet (Google): Q1 2026
- Microsoft: FY2026 Q3
Markets will be watching: (1) whether AI capex is justified by real growth (Azure at 39% YoY in Q2, Google Cloud projected above 50% in Q1), (2) the impact of Trump’s tariffs on guidance, (3) whether anyone can monetize AI beyond cloud ARR.
April 30 — Apple Q2 FY2026. First post-tariff test: Morgan Stanley estimates $33 billion annualized in additional costs. The tariff pause that enabled the 600-ton iPhone airlift from India is temporary.
Hexagon/Waygate: the regulatory process for the $1.45 billion deal begins (close H2 2026).
Thanks for reading.