Saturday, February 28, 2026 · 65 newsletters

The Guardrails Bend, Then Break

saaspocalypse · anthropic-vs-openai · epstein-as-system · iran-clock · tariff-architecture · ai-labor-narrative · election-machinery · media-consolidation

February began with a $357B Microsoft session that finally gave the AI capex bubble a price tag, and it ended with the United States at war with Iran without an AUMF, a Supreme Court that had stripped the legal basis for the administration's tariff regime, and Anthropic in court over a Pentagon supply-chain designation. The legal, constitutional, and commercial guardrails that held the second Trump term in check for thirteen months all bent or broke in the same four weeks. The arcs that began as separate stories in week one converged into a single story by week four.

The Month in One Sentence

This was the month the software business model got repriced in public, the AI labs picked sides on whether to be commandeered by the state, and the Republic cracked in three independent places at once.

Arc: The SaaSpocalypse, From Tremor to Trillion-Dollar Repricing

Week one read like a market correction. When Microsoft shed $357B in a single session after Q2 earnings, Runtime's Tom Krazit called it "A day of reckoning for the AI boom," and Ben Thompson at Stratechery ran the contrarian frame for why Meta rallied on similar capex. The reading was still narrow: a Big Tech earnings story with a side of Anthropic raising at $350B.

Week two named the regime. Anthropic shipped Claude Opus 4.6 with a one-million-token context window and a legal-work tool, and the software ETF was down nearly $1T inside seven days. Jefferies trader Jeffrey Favuzza coined the line that stuck. Noah Smith called it "The Fall of the Nerds," and Ben Thompson followed with "SaaSmageddon," clocking Thomson Reuters down 21% on its biggest single-day move ever, Gartner off 21%, RELX off 14%. Linas Beliūnas calculated $285B wiped out in four trading days from a single Anthropic plug-in announcement. The capex on the other side hit roughly $650B across four hyperscalers, more than the US interstate highway system.

Week three turned it into a sector test. Madison Mills at Axios AI+ put $400B on the table after the Anthropic Super Bowl spot, and the rout spread beyond software. CBRE and JLL down 12%, Cushman down 14%. By Thursday, Bloomberg had CH Robinson down 24% intraday after a white paper from a former karaoke company named Algorhythm Holdings claiming 300 to 400% volume scaling without headcount. The repricing was no longer narrative. It was tape.

Week four made it structural. The Average Joe's "Klarna's rough patch" had Klaviyo, Sprout Social, and HubSpot down as much as 75% over twelve months. App Economy Insights framed it cleanly in "Seats vs. Compute": vendors who sell human productivity get punished, those capturing the rise in workloads get rewarded. By the final week, Greg Isenberg's "Skills Eat SaaS" had named the unit of value shift directly: "SaaS packages predictable results. Skills package judgment." And Trung Phan at SatPost via Contrary Research had the running tally at over $1T in lost B2B SaaS market cap year to date. What started as a Microsoft earnings story ended as the largest sector repricing of the AI era.

Arc: Anthropic Pulls Away, Then Picks a Fight With the State

Week one was the quiet coronation. John Ellis at News Items led with Mark Gurman's Friday scoop that Apple internal development "runs on Anthropic at this point," with Apple wanting to rebuild Siri around Claude. Anthropic doubled its fundraising target from $10B to $20B at a $350B implied valuation. Linas Beliūnas wrote that Anthropic's MCP Apps launch with Amplitude, Asana, Box, Canva, Figma, Hex, monday.com, and Slack as partners marked the moment ten sophisticated SaaS companies handed their user relationships to an AI company.

Week two made it commercial warfare. Anthropic spent millions on four Super Bowl spots opening with "BETRAYAL," "TREACHERY," "VIOLATION," and "DECEPTION" cards, with an actor resembling Sam Altman recommending insoles. Tech Brew had Anthropic spending $14M on game night, part of an AI industry dropping $382.8M on traditional TV. Trung Phan at SatPost framed it as "Drake vs. Kendrick" for the AI era, with the structural point that Anthropic earns 80% of revenue from enterprise and can afford to torch the consumer narrative. By Sunday the FT had Anthropic guiding investors to over $30B annualized revenue by year end.

Week three landed the financial mark. Techmeme led Thursday with Anthropic closing a $30B Series G at a $380B post-money, run-rate revenue at $14B and Claude Code alone at $2.5B. By Sunday the Pentagon had told Axios it might sever the Anthropic relationship over AI safeguards, after Anthropic asked questions about how Claude was used in the Maduro raid. Casey Newton at Platformer framed it as the moment AI safety researchers had warned about for years.

Week four was the break. On Tuesday the Pentagon set a Friday deadline: grant unfettered Claude access including bulk surveillance of Americans and autonomous lethal weapons, or be designated "a supply chain risk," a penalty usually reserved for foreign adversaries. By Friday, Anthropic had refused. Ross Andersen at The Atlantic had the scoop on what broke the talks: the Pentagon wanted Claude to "analyze bulk data collected about Americans." On the same day, Newcomer led with OpenAI closing $110B from Amazon, Nvidia, and SoftBank at an $840B post-money. The WSJ then reported the Pentagon used Claude in its Iran air attack hours after Trump declared the federal government would end its use of Anthropic. One lab chose the contract. The other chose the courtroom. That divergence will price into recruiting and procurement for years.

Arc: Epstein Stopped Being a Scandal and Became an Operating System

Week one was the bipartisan moment. Three million pages dropped Friday before the month began. Gov Brief Today flagged the operational disaster: DOJ published dozens of unredacted nude photos of young women while remembering to redact Trump's face. The Times found 38,000 references to Trump, his wife, and Mar-a-Lago. The reading was still discrete: a scandal with a base-versus-elites frame.

Week two routed the story through private equity. Semafor Business made the smartest observation: Brad Karp losing his seat atop Paul Weiss was not really an Epstein story, it was a story about overexposure to PE. Karp's relationship with Apollo's Leon Black was what pulled him into the orbit. Semafor's punchline: the American economy now looks like Big Law, "overtorqued toward financial firms." By Friday the cascade reached London, with Latika Bourke reporting Epstein wired Peter Mandelson $75,000 in 2003.

Week three hit the cabinet. Popular Information's Judd Legum had the cleanest reporting: Commerce Secretary Howard Lutnick's "one and absolutely done" story was demonstrably false. He emailed Epstein in December 2012, four years after the first guilty plea, to arrange a get-together on Little St. James. By Tuesday Lutnick admitted under questioning that he stayed in contact through 2018. The polling shifted underneath it: Americans now said Biden did a better job by 46 to 40 per YouGov, and Trump slid from +6 to -14 approval.

Week four moved it to systems. Matt at Crooked Media's "Shame of Thrones" framed it: "British police arrested a royal for the first time in almost half a millennium." Former Prince Andrew was arrested on his 66th birthday. Anand Giridharadas at The Ink launched "The Epstein Class," a series arguing that what we need to understand is the operating system of elite power, not just the names. The Bannon-Epstein texts surfaced, revealing Bannon called Epstein "God" and spent months in 2019 advising him on how to rebuild his image. The story stopped being a base-versus-elites distraction and became a polling regime change, a cabinet-credibility problem, and a transatlantic political crisis in three news cycles.

Arc: The Coalition Math on Immigration Inverts in Twelve Months

Week one was the Minneapolis floor. ICU nurse Alex Pretti was shot in the back by federal agents, and for the first time in the second term, evidence visibly moved actors who do not usually move. Republican governors broke ranks. JVL at The Bulwark called it "Our Gettysburg Moment." Minneapolis held the first citywide general strike in America in decades. Anand Giridharadas at The Ink covered the scale. ICE withdrew 700 agents and pulled Greg Bovino back to El Centro.

Week two added the Don Lemon arrest. Federal agents arrested former CNN anchor Don Lemon and local journalist Georgia Fort for filming a protest. The detail that mattered: a magistrate, an appellate judge, and career prosecutors all refused the case before Pam Bondi got her indictment. JVL framed it as "a flashing, neon red-line being crossed." When language converged across siloed publications that quickly, the consensus narrative had shifted.

Week three flipped the polling. Dan Pfeiffer at The Message Box cited Nate Silver averages showing Trump 12 points underwater on immigration. Jim Swift at The Bulwark flagged the converging number: fewer than 14% of the 400,000-plus immigrants ICE has arrested have been charged with or convicted of a violent crime. By Friday night DHS funding lapsed. The party that won 2024 on the issue lost the issue inside a single news cycle.

Week four moved the machinery. Democracy Docket obtained an April 2025 draft executive order that would allow Trump to unilaterally ban mail-in ballots and voting machines on a national-emergency theory. Adrian Carrasquillo at The Bulwark had DHS official Heather Honey, an architect of 2020 stolen-election claims, promising election officials ICE would not be at the polls in November. Arizona's Adrian Fontes and Maine's Shenna Bellows both said they did not believe her. The coalition math moved roughly 30 points in twelve months.

Arc: The Tariff Architecture, From Inflation Story to Constitutional Rebuke

Week one was a Kevin Warsh story. Bloomberg called Trump's pick "curious." Paul Krugman titled his take "A Bad Heir Day at the Fed." Matt Stoller's Boomcession framing named the regime under the politics: wage growth up, consumer confidence at a 12-year low. The macro had a name; the trade policy did not yet have a court ruling.

Week two cited the receipts. Paul Krugman demolished the Bessent and Miran argument that AI productivity gains justified Fed cuts. 52% of Americans said Trump had made the economy worse. By Thursday, Semafor Business called it "Trump's two-track stimulus": One Big Beautiful Bill tax cuts plus a backdoor $200B MBS purchase via Fannie and Freddie. Jamie Dimon called it QE at Davos.

Week three named the magnitude. Brian Daitzman at Lincoln Square argued the 2025 tariff regime was the largest US tax increase since 1993 as a share of GDP: effective rate at 9.9%, import-weighted applied rate at 13.5%, highest in the postwar era. Average household cost about $1,000 last year. Krugman walked through new CBO and NY Fed studies confirming tariffs were borne by US households and firms.

Week four cracked the legal basis. On Friday, SCOTUS in Learning Resources v. Trump struck down the IEEPA tariffs 6-3, with two Trump appointees joining Chief Justice Roberts. Trump pivoted to a 10% Section 122 tariff within hours, but Matt Klein at The Overshoot noted Section 122 only lasts 150 days and the original collections still need to be refunded. Penn Wharton estimated $500M in IEEPA tariff revenue had been collected daily. By the end of the month, the refund number had landed at $130B per Dexter Roberts at Trade War, and Justice Kavanaugh had openly described the unwind as a legal "mess." A month that started with a Fed chair pick ended with the executive's signature economic tool stripped by the court.

Arc: The Agent Stack Hardens, the Skeptics Get a Canon

Week one was Moltbook. Peter Steinberger's social network for AI agents accumulated 30,000-plus agents being browsed by 3,000 humans at any moment. Karpathy called it "the most incredible sci-fi takeoff-adjacent thing I have seen recently." Casey Newton at Platformer published "Falling in and out of love with Moltbot," the cleanest practitioner report.

Week two was the org chart. Dan Shipper at Every ran "The Two-Slice Team," arguing Opus 4.6 and Codex 5.3 collapsed the two-pizza rule. Every ran four software products with one person each, 99% of code written by AI agents. Lenny Rachitsky's interview with Sherwin Wu had 95% of OpenAI's API engineers running 10 to 20 parallel Codex agents. Nikhil Basu Trivedi reported that 4% of GitHub public commits were now authored by Claude Code, projected to hit 20% by year end.

Week three was the runtime. Nikunj Kothari's "Token Anxiety" became the canonical 2026 image: dinner conversations shifting from "what are you building" to "how many agents do you have running," a friend leaving a party at 9:30 on a Saturday to get back to his agents. Work-Bench named the stack: execute, constrain, observe, improve. G2 had 57% of companies with AI agents in production.

Week four was the skeptics catching up. Ruben Hassid's "Workaholic" cited a Harvard 8-month study at a 200-person tech company where everyone chose to use AI and everyone ended up working more. PMs wrote code, researchers did engineering. AI removed the friction of starting and stopping; if stopping was harder than continuing, you had an addiction. Trung Phan at SatPost proposed the cleanest signal of the month: AGI has arrived when all the major AI labs stop running on Slack. Amazon's internal AI coding assistant caused at least two AWS outages including a 13-hour disruption, after the bot decided the engineers' existing code was inadequate and deleted it. Block laid off 40% citing AI; the stock jumped 24% before Jack Dorsey admitted on X the real issue was COVID-era overhiring. Om Malik called it "narrative substitution." The infrastructure got real, the prices dropped, and the literature about what these tools do to us finally caught up with the literature about what they can do.

The Story of the Month

The story of February 2026 is not the Iran strikes, the IEEPA ruling, the Lutnick admission, or the SaaSpocalypse on their own. It is that all of them ran on a single rising clock, and by the final week they fed each other. The Pentagon's escalation against Anthropic landed in the same eight days as a regime-change war launched without an AUMF, a Supreme Court that wiped out the legal basis for the administration's tariff regime, a White House that pivoted to a separately illegal statute within days, and a 17-page draft executive order to ban mail-in ballots on a national-emergency theory. The same week Michael Flynn convened a room of election deniers to plan exactly that, the WSJ confirmed the Pentagon had used Claude in the Iran air campaign hours after Trump declared the federal government would stop using Anthropic.

The case for treating this as the story rather than five parallel stories is that the velocity is the news. Each individual rupture would be the political story of any other year. They all happened in the same four weeks, and the AI capital cycle, the Epstein cascade, the immigration polling inversion, the tariff legal architecture, and the war powers question stopped running on parallel tracks and started running on one track. Any read that filed them in separate genres missed the load-bearing pattern: the legal, constitutional, and commercial guardrails that held the second Trump term in check for thirteen months all bent or broke in the same month.

In Retrospect

The "feels like 2018" framing aged poorly within days. Semafor DC's domestic-challenges-mount-for-Trump take in week two read accurately at the time. By week four, Trump had launched a regime-change war, taken OpenAI's $110B side, and had a draft executive order to seize the midterms in circulation. 2018 was a year of bad polling; 2026 became something else inside two weeks of that filing.

The SaaSpocalypse was filed as a market correction in week one. The framing was a Microsoft earnings story, and analysts called for a rebound. By month end, $1T in B2B SaaS market cap had vanished year to date, JPMorgan was titling notes "Software Collapse Broadens with Nowhere to Hide," and Goldman had rolled out a new long-short basket to navigate the new regime. The category itself was being rewritten, not corrected.

The Anthropic-Pentagon standoff read as a procurement squabble in week three. Most newsletters covered the Sunday Axios scoop as a Hegseth news cycle. By Friday of week four, the Pentagon was using the Defense Production Act framework, threatening a designation reserved for foreign adversaries, and Anthropic was preparing to sue. The frame that aged worst was treating "AI safety" as a values discussion rather than a legal and contractual one. By Saturday it was the contractual one.

The "AI is eating jobs" narrative held through three weeks. Block's 40% cut with a 24% stock pop on a Friday looked like the cleanest case study yet. Hours later Dorsey was on X admitting the actual problem was operational. Om Malik called it AI as cover for ZIRP-era rot. The Oxford Economics report and the IBM survey both landed in the same news cycle: 84% of execs called AI transformative, 26% used it effectively, 5% had agentic success metrics, and 80% of companies reported no productivity gain. The labor story was load-bearing for cover stories that had nothing to do with productivity.

What to Carry Into Next Month

The two clocks framing has hardened. The visible clock this month was Iran, the IEEPA ruling, the SOTU, the Super Bowl, the cabinet hearings. The faster, quieter clock was the redistricting math, the FBI's mystery midterm call, the Flynn dinner with Cleta Mitchell, the DOJ suing five more states for voter rolls, the IRS feeding taxpayer addresses to ICE in 42,695 documented violations. By the final week, the quieter clock was visibly doing more of the structural work. The Pfeiffer-Beutler split on whether the Resistance has structural momentum or just one good week is the most important strategic argument inside the Democratic party right now, and treating any of February's individual ruptures as a discrete crisis is the misread that will look very foolish by November.

The AI conversation has split into two coherent literatures that no longer pretend to be one conversation. The first is the operator stack: Sonnet 4.6 at one-fifth Opus pricing, Google's three-protocol commerce OS, Stripe's HTTP 402 rail, Mistral at $400M ARR, Every running five always-on bots in a shared Discord. The second is the skeptical canon: the Harvard work-hours study, the Slack AGI Test, the Amazon coding bot deleting engineer code, the Block AI-cover narrative, the 80% of companies with no productivity gain. Both are real, both are growing, and the operators who treat them as opposing camps are missing that the runtime is hardening and the cost is being named at the same time. The question is no longer whether AI works in production. It is who captures the rent and what kind of person AI work makes you.

The reading list to carry forward is short. Edwin Eisendrath's "While We Slept, Donald Trump Launched a War" is the cleanest read on the AUMF and the constitutional stakes, and it will be re-read in retrospect as the moment the war powers question stopped being theoretical. Ross Andersen at The Atlantic on the Anthropic-Pentagon dispute is the inside story of what actually broke the deal, and it is the document that will define how the two lab paths get taught. Om Malik's "Block & Tackle: Job Cuts & the AI Narrative" is the sharpest piece on how the AI labor story is being used as cover, and it is the frame to bring to every executive cost-cutting announcement for the rest of 2026. February told three things in sequence: the war powers are now whatever the executive says they are, the AI labs are now state-aligned or they are not, and the software business model that built the last twenty years was repriced in public. Those are the three frames to carry into March.