April 02, 2026
Matt Levine (Bloomberg) · rss · 16 mins
OpenAI just closed a $122 billion private funding round at an $852 billion valuation — more money than SpaceX’s still-hypothetical IPO (projected at up to $75 billion, which would already dwarf Saudi Aramco’s record $29 billion 2019 debut). CFO Sarah Friar said the round “blows out of the water even the largest IPO that’s ever been done.” Over $3 billion came from retail-adjacent channels: bank private wealth desks and Cathie Wood’s ARK Invest ETFs.
Private markets have structurally replaced public markets for mega-fundraising. OpenAI raised $122 billion while still private; it can also sell stock to quasi-retail investors via bank wealth channels and ETFs. The traditional IPO logic — go public so the broader investor base can buy in, driving up your price — collapses when you’ve already sold to everyone who wants in at a private-market premium.
Despite the record raise, OpenAI shares have quietly collapsed on the secondary market. About half a dozen institutional holders (hedge funds and VCs) approached Next Round Capital with ~$600 million of OpenAI stock to sell — shares that “would have been snatched up within days” a year ago. Now no buyers. Meanwhile, secondary platforms report “$2 billion of cash ready to deploy into Anthropic,” with buyers citing better risk/reward at Anthropic’s $380 billion valuation vs. OpenAI’s $852 billion.
Snap’s non-voting share structure — public shareholders have literally zero votes — makes it an extreme test case for influence-without-power activism. Irenic Capital, with a 2.5% stake, published a letter demanding layoffs and spinning off Specs, noting Snap has underperformed the Nasdaq by 444 percentage points since its 2017 IPO (a $1 investment is now worth $0.23). The stock jumped 13% on the letter anyway, suggesting embarrassment and peer pressure do exert real pressure even without proxy-fight leverage.
“Floating-strike convertibles” (sometimes called death-spiral convertibles) are the financial instrument that lets distressed micro-cap companies monetize their lottery ticket of becoming a meme stock. The mechanics: a hedge fund gives the company $100 now; later it can convert at a lookback price — e.g., the lowest daily price over the prior 10 days — so it profits enormously if the stock briefly spikes. If no meme moment comes, the fund converts steadily at market prices, selling into the stock and pushing it lower in a self-reinforcing spiral.
Algorhythm Holdings — a publicly-listed former karaoke company — made bold AI-for-trucking claims last month, briefly erasing ~$17 billion from real trucking stocks while its own stock spiked 450%. Chicago investor John Fife’s Streeterville Capital had previously bought floating-strike convertibles in Algorhythm and sold millions of shares into the spike for potentially millions in profit. The SEC had previously called Fife a “recidivist violator,” noting he had targeted ~135 microcaps in sectors including “marijuana, blockchain, bitcoin, vapour, lithium and gold mining.”
Ajinomoto — the Japanese company that invented MSG and discovered umami — turns out to control a critical bottleneck in AI chip packaging. Its Ajinomoto Build-up Film (ABF), a high-performance insulating polymer developed from food-science chemistry, is the industry-standard insulating film for high-end semiconductor substrates with no qualified substitutes. Activist Palliser Capital argues the ABF business is hidden inside a “Healthcare & Others” segment, creating a 70%+ valuation discount vs. its own customers, and is calling for a 30%+ ABF price increase.
Anthropic accidentally published source code for its Claude coding agent, revealing that Claude detects user frustration via a hardcoded regex — scanning for terms like “wtf,” “ffs,” “this sucks,” “piece of [expletive],” “broken,” “useless,” etc. An LLM using regex instead of inference for sentiment is peak irony, but it’s also correct engineering: a regex is faster and cheaper than an LLM call just to check whether someone is swearing at your product.
Quotable:
“We literally couldn’t find anyone in our pool of hundreds of institutional investors to take these shares.” — Ken Smythe, Next Round Capital, on ~$600M of OpenAI stock offered for sale on the secondary market with zero takers
Ben Thompson · rss · 8 mins
On March 30, 2026, attackers compromised the npm account of @jasonsaayman, a lead axios maintainer, via social engineering — seizing a long-lived classic token that survived npm’s post-Shai-Hulud credential reforms. Within 39 minutes they published poisoned versions across both the 1.x and 0.x release branches of axios, the JavaScript HTTP client with over 100 million weekly downloads. The malicious payload wasn’t in the axios source code itself but in a phantom dependency, plain-crypto-js@4.2.1, which ran a postinstall script deploying a cross-platform RAT on macOS, Windows, and Linux, then self-destructed and swapped in a clean package.json to frustrate forensics.
The attack was carefully staged: 18 hours before the axios releases, the attacker published a clean version of plain-crypto-js under a separate account to build publishing history and evade new-package scanners. This is the third major npm supply chain attack in roughly a year — after the Shai-Hulud worm (September 2025, 500+ packages compromised, CISA advisory issued) and Koi Security’s PackageGate research (January 2026, six zero-days across npm/pnpm/vlt/Bun). Every post-Shai-Hulud hardening measure — OIDC trusted publishing, FIDO 2FA, 7-day token caps, SLSA — was bypassed because the reforms hardened everything downstream of the maintainer account but left the account credential itself as a single point of failure.
Separately, Anthropic accidentally published a 59.8 MB JavaScript source map file in Claude Code npm package version 2.1.88, exposing the internal source of its agentic harness. Claude Code is at $2.5 billion ARR (more than doubled since January 2026), with enterprise clients representing 80% of revenue. Anthropic confirmed the leak was “human error, not a security breach” with no customer data or credentials exposed — and notably, Claude Code itself depends on axios, meaning it was also affected by the supply chain attack (which is an actual security breach, unlike the source leak).
The leaked code is copyrighted and can’t be legally copied, but the strategic damage is real: competitors like Cursor and Microsoft building their own harnesses on Claude models gain the most insight, not OpenAI. The harness alone isn’t the full moat — the deeper differentiator is the integration of harness and model — but the leak still represents a meaningful IP exposure for a company at that commercial velocity.
Both incidents together illustrate a broader security thesis: AI will make security worse before it makes it better. Near-term, vibe-coding increases bug surface and AI tools help malicious actors find and exploit vulnerabilities faster. Long-term, AI is the only realistic solution — it can audit entire codebases (not just new code), trace full dependency trees no human would inspect, execute inconvenient-but-secure workflows without fatigue, and stress-test security flows repeatedly at scale. The axios attack specifically demonstrates the core human failure modes: convenience-seeking (lazy dependency hygiene), long-lived credentials left unrevoked, and social engineering exploiting human trust.
Quotable:
“AI can navigate an extremely inconvenient but highly secure workflow, and can stress test every aspect of that flow — repeatedly — in a way no human can.” — on why AI is the long-term answer to security, despite making things worse in the short term
Bloomberg · email · 10 mins
US private equity has identified Japan’s $227 billion restaurant market as undervalued and stalled — sector growth has flatlined since 2018 despite structural tailwinds. Rising inflation, a surge in one-person households, and time-poor dual-income families are driving demand for affordable, convenient food, creating an opening for chains with scale and brand recognition.
Carlyle Group paid ¥135 billion ($847 million) for KFC Japan in 2024, and Goldman Sachs’s PE arm paid ¥70 billion for Burger King Japan. Both are now executing aggressive expansion playbooks: Carlyle plans to grow KFC’s ~1,300 outlets by 30% to ~1,700 within four years; Goldman Sachs aims to nearly double Burger King Japan’s ~350 shops to 600 by 2028.
KFC Japan holds a cultural moat that goes beyond menu: a 1974 marketing campaign established fried chicken as a Christmas tradition, with queues “snaking out the door” each December. That embedded ritual gives KFC brand loyalty that no new entrant can easily replicate — and PE sees it as untapped upside, not just a legacy asset.
Both KFC and Burger King Japan are leaving daypart revenue on the table. Carlyle’s Takaomi Tomioka, who led the KFC acquisition, notes KFC has underserved breakfast, snacks, and late dinners compared to McDonald’s, which operates more than 3,000 locations in Japan. The operational playbook includes self-order kiosks to cut labor costs, expanded delivery, consumer data analytics, and localized premium menu items.
McDonald’s and Mos Burger (Japan’s homegrown No. 2 with just over 1,300 locations) are already showing higher average spend per visit even as they raise prices — validating the thesis that Japanese consumers will trade up within fast food rather than exit the category. Wendy’s Japan, chaired by former Domino’s Japan head Ernie Higa, is now fielding acquisition offers, signaling the deal wave is not over.
Quotable:
“Food can be globalized. But at the same time, it needs to be local.” — Takaomi Tomioka, Carlyle executive who led the KFC Japan acquisition
‘Lenny’s Newsletter’ via PubsforSubs · email · 10 mins
The rules of organizational life have fundamentally changed. Organizations are flattening, layoffs are structural (not just cyclical), and AI is enabling fewer people to do more — all of which compress scope and resources for ICs and frontline managers while executives drive these changes rather than absorb them. Great people end up stuck with overwhelmed managers who lack the bandwidth to advocate for them, resulting in blocked promotions and wrenched-away projects for structural, not performance, reasons.
The core stay-or-go framework: stay if the role keeps you current, leave if it lets you fall behind. The way product gets built is transforming every six months via AI tooling, flatter orgs, and wider individual scope — a household-name company that refuses to transform is now becoming less valuable on your resume, not more. Short tenures are more accepted than most people assume; staying in the past is more damaging than making a quick jump.
High comp is the one valid reason to stay in a bad situation, but only if you use the time wisely. A stay bonus or stock run-up pushing comp near $1M can fund the savings and runway needed to later take risk at a faster-moving company. The tradeoff: you must actively work to stay current — champion AI-forward projects internally, find a community or side hustle that keeps your skills sharp — otherwise comp buys time but the gap compounds.
Don’t put people on the spot until you’ve earned the right. Confronting a manager about job security during a reorg backs them into a corner — they can’t reassure you without lying. Going to a skip without a heads-up breaks trust they may never restore. A staff engineer in one case listed promotion, visibility, and scope expansion as goals on day one with a new manager — the problem isn’t the goals, it’s that stating them before delivering anything makes you look out of touch when your first review is anything less than stellar. Let competence create the leverage first, then have the conversation.
When things feel unstable, the best 30-day play is unglamorous: put your head down and deliver. In 50/50 decisions about who stays and who goes, recent strong execution tips the scale. Expressing anxiety primes your manager to perceive you as disengaged even when your performance says otherwise — anxiety is contagious. If you’re confident a change will force you out, get ahead of it internally (companies remove entire teams, so early awareness helps you land a role before the music stops), but there’s little advantage to jumping to external interviews prematurely.
Cross-functional relationships are an insurance policy that must be built before the rainy day, not during it. People who get advance intelligence on reorgs aren’t the ones who start asking the right questions when things go dark — they’re the ones who already have a peer they can pull aside. Investing only in your manager and skip creates a single point of failure. Building these relationships requires generosity without a transaction in mind; trust cannot be manufactured on the spot when you need it most.
When the promotion contract breaks — as it did for a manager who had VP sponsorship and calibration approval for a director promotion but was still blocked because leadership doesn’t want to create roles that may not exist in two years — the burnout that follows is about the gap between what you gave up and what you got back, not workload per se. The prescription: take 10-15% off the gas and test whether performance holds; use AI tools to reclaim hours; cut the bottom 5-10% of calendar. Protecting what matters outside of work (health, family, relationships) is what makes the work sustainable and prevents the malice that drives people out of the industry entirely.
Quotable:
“Directness without leverage is just pressure — and pressure makes people close down, not open up.” — on why confronting managers or skips about org drama before you’ve earned trust backfires
The Core · email · 8 mins
When Novo Nordisk’s Indian semaglutide patent expired on March 20, 2026, sixteen generic versions of Ozempic hit Indian pharmacy shelves within 48 hours and prices crashed 70–90%. Zydus launched three brands and anchored API supply for Lupin and Torrent; Dr. Reddy’s launched “Obeda” with plans for 87 countries; Torrent became the first Indian company to offer a generic oral tablet alongside the injectable. No country on earth executes the post-patent rush faster.
The scale of India’s copying machine is enormous — it supplies 40% of America’s generic medicines, two-thirds of the world’s vaccines, and clocked Rs 2.60 lakh crore in pharma exports in FY25 (total industry turnover Rs 4.72 lakh crore) — yet earns zero on originator economics. The royalties, licensing income, and durable pricing power all flow to whoever invented the molecule. Every patent expiry triggers a race to the bottom: fifty brands enter, margins collapse within weeks, and the comfortable returns that funded three decades of generics growth get thinner each cycle.
Cipla chairman Yusuf Hamied made the structural bet explicit in a 2015 Lancet interview: “Our R&D at Cipla is targeted at incremental innovation — how to change and improve a product that already exists.” India’s drug companies built their entire R&D apparatus around extending and improving molecules someone else discovered — a rational choice when copying was faster, cheaper, and nearly guaranteed to succeed, but one that locked the industry out of first-mover rents entirely.
The odds against discovery are brutal and well-documented: a new drug takes 10–15 years, costs $1–2 billion, and fails nine times out of ten. Kallam Anji Reddy understood this and tried anyway — he launched India’s first drug discovery programme in 1993 alongside the generics business making him rich. The generics side was spectacular (Dr. Reddy’s became the first Indian company to file for US drug approval and earned Rs 366 crore in FY2002 from a single 180-day exclusivity on generic fluoxetine 40mg); the original-drug side did not produce a first-in-class molecule.
The current wave of Indian drug discovery programs is the most serious since Dr. Reddy’s was licensing molecules to Novo Nordisk in the 1990s. Among three active obesity drug candidates, Zydus is furthest along. Whether this generation can clear the bar that eluded every previous attempt — inventing a novel molecule for global markets — remains the central unanswered question for the world’s largest generic drug industry.
Quotable:
“Our R&D at Cipla is targeted at incremental innovation — how to change and improve a product that already exists.” — Yusuf Hamied, Cipla chairman, in a 2015 Lancet interview
Capital Gains · email · 6 mins
Market data is structurally adversarial in a way AI training data is not. A Reddit thread about dying lawn grass yields dozens of earnest, truthful answers — people genuinely trying to communicate. A trading signal is the opposite: it’s evidence that some other participant is making an error, and every participant is simultaneously trying to hide their own edge and exploit others’. This is why clean data beats more data in AI research but means something different in finance: the market’s “data” is a record of deliberate misdirection, making it closer to training a model on North Korean propaganda than on Wikipedia.
Successful signals erase themselves. Suppose sloppy large-sell orders reliably push a stock down 1% intraday. A trader who spots this buys at -0.5%, capturing half the move. Now the data shows only a -0.5% dip on heavy volume — not a 1% drop. A second trader, seeing the smaller signal, calibrates to buy at -0.45 bps, shrinking it further. Eventually a careful analyst studying the data reaches a “null result”: equal-sized orders at equal intervals no longer move prices. They aren’t noise — they’re already someone else’s captured signal, invisible to anyone who didn’t already know to look.
Correlated crowding creates synchronized blowups, and market stability is itself a leading indicator of fragility. The Quant Quake of 2007 is the canonical case: funds “diversified” across fixed income arbitrage, equity statistical arbitrage, value, and momentum were actually all running similar leveraged strategies. When one fund began shrinking its balance sheet, all of them did simultaneously — everything went down in unison. The deeper problem is structural: when markets seem solved and volatility is low, the rational response is to lever up, which guarantees the next crash. The system feels most perfectly understood right before a 1987-style break or a 2008-sized collapse, and that correlation-going-to-1 is a test you only get to run once.
Quotable:
“It’s like training a language model exclusively on North Korean propaganda and then asking it to write a coherent story of the twentieth century: much of what you’d want to train on is simply missing from the data.” — on why market signals are structurally absent from the historical record once they’ve been exploited
FT Newswrap · email · 6 mins
The Iran war — the second energy shock in under five years — is reviving nuclear power globally. Fatih Birol, IEA head, called it “the greatest global energy security threat in history,” comparable to the 1973 and 1979 oil shocks, which themselves produced over 40% of today’s existing nuclear capacity. The crisis is now catalysing a similar policy wave.
Germany, long the most committed anti-nuclear country in Europe, is cracking. Energy minister Katherina Reiche stated there is “no alternative” to gas now that previous governments shut down the country’s reactors — while ruling out restarting conventional plants, Germany is backing small modular reactors and has pledged to stop blocking nuclear at the EU level. Sweden, Poland, France, the UK, Japan, and South Korea are all expanding or extending existing nuclear capacity.
The Chernobyl containment structure — a massive steel shell built to safely seal the world’s worst nuclear accident — was struck by a Russian drone and is now in a precarious state, keeping alive the safety fears that slowed nuclear adoption after Fukushima in 2011. Nuclear fusion, once impractical, is also attracting investor interest, with the UK and Germany among those pursuing it as a long-term clean energy source.
Quotable:
“The greatest global energy security threat in history.” — Fatih Birol, IEA head, on the Iran war’s impact on global energy markets
Cory Weinberg · email · 6 mins
OpenAI’s $122 billion raise at an $852 billion post-money valuation looks strong on the headline but is murkier underneath. The $12 billion tranche from financial investors was 20% oversubscribed, but the biggest checks came from existing and strategic investors — Amazon and Nvidia — who have customer relationships with OpenAI and aren’t purely seeking financial returns.
Secondary market data from Caplight, which aggregates transaction data from hundreds of SEC-registered broker-dealers, shows $1 billion in sell orders versus only $200 million in buy orders for OpenAI shares in 2026 through March. CEO Javier Avalos notes the typical seller is offloading $50M+ stakes of preferred stock, meaning institutional investors — not retail — are seeking exits. This is a “huge reversal” from Q3 and Q4 2025, when demand dominated and supply was minimal.
OpenAI’s secondary selling pressure far exceeds that of Anthropic and SpaceX, both of which are also approaching IPOs. SoftBank, which holds roughly 25% of its total asset value in OpenAI and is the best public proxy for OpenAI’s implied valuation, is down 17% in 2026 despite CEO Masayoshi Son increasing his stake. OpenAI dismisses Caplight’s data as “not reflective of demand through authorized channels,” but the IPO will settle the debate — and Altman’s investors may find buyers at a lower valuation than $852 billion.
Quotable:
“This is a huge reversal from Q3 and Q4 of 2025, when we saw mostly demand in the market and minimal supply.” — Caplight CEO Javier Avalos on the secondary market shift for OpenAI shares