Per-Seat Software Built a $390 Billion Industry. AI Agents Just Broke the Math.

On the morning of February 3, Jeffrey Favuzza at Jefferies' equity trading desk watched Goldman Sachs' basket of U.S. software stocks drop 6% in a single session. It was the biggest one-day decline since April's tariff selloff. Bloomberg's headline captured the mood on trading floors from Manhattan to Mumbai. "Get Me Out."

Favuzza coined a term for it. SaaSpocalypse. The word stuck because the selling didn't stop.

Over 48 hours, $285 billion evaporated from global software stocks. Thomson Reuters posted its largest intraday decline in company history, falling 17%. LegalZoom lost a fifth of its value. Monday.com, the project management platform that had become a Wall Street darling, was down 50% for the year by the time it reported earnings on February 9. Even companies that beat estimates got punished. The market wasn't pricing in bad quarters. It was pricing in a broken business model.

The trigger was Anthropic. On January 12, the AI company launched Claude Cowork, a general-purpose AI agent that can read, modify, and create files on a user's computer. On January 30, it released 11 open-source plugins covering legal, sales, finance, marketing, and customer support. On February 5, it shipped Claude Opus 4.6 with coordinated agent teams and a million-token context window. Not a research paper. Not a demo at a developer conference. A shipping product that reviews contracts, manages CRMs, and processes invoices, available to anyone paying $100 to $200 per month.

BTIG's chief market technician ran the numbers and found that software stocks had underperformed semiconductors by 20 percentage points over 20 trading days. The last time anyone saw a divergence that extreme was right before the dot-com floor gave out in early 2000.

The question nobody on Wall Street could answer cleanly was whether this was panic or math. A $390 billion global industry built on per-seat licensing just watched an AI company demonstrate, in production, that software agents can do much of what those seats were paying humans to do. If the agents work, the seats don't make sense. And if the seats don't make sense, neither do the revenue multiples.

The Rout, Day by Day

The selling had a structure to it. Each wave corresponded to a specific Anthropic release hitting a specific category of software.

Cowork's initial launch on January 12 rattled startups but didn't move public markets much. It was a research preview, macOS only, limited to Claude Max subscribers. The January 30 plugin release changed the calculus. Eleven plugins, all open-source, covering the exact workflow categories that public SaaS companies charge per-seat premiums for. The legal plugin alone could automate contract reviews, NDA triage, and compliance workflows. That plugin is what sent Thomson Reuters and RELX into freefall.

February 3 was the worst day. Goldman's software basket fell 6%. The financial services index dropped nearly 7%. Private credit stocks, exposed through loans to software companies acquired at peak 2022 and 2023 valuations, fell in sympathy. Ares Management and KKR both lost 10%.

February 4 hit harder in specific sectors. Thomson Reuters fell 17%, its worst session ever recorded. RELX, the parent of LexisNexis, matched it. Wolters Kluwer dropped 13%. India's Nifty IT Index lost nearly 6% as Infosys posted its steepest single-day decline in more than two years. Roughly $24 billion in market capitalization disappeared from Indian IT stocks in a single session.

February 5 brought Opus 4.6 and the agent teams feature. The IGV ETF, which tracks major software companies, hit its 52-week low at $79.65. It had been falling every single trading day for over a week.

By February 9, Monday.com reported earnings that beat analyst estimates, then issued weak 2026 guidance. The stock fell another 17 to 20 percent. The message was clear. Beating the quarter didn't matter if the model was broken.

Not everyone got hit. Palantir rose 11% after reporting 137% growth in U.S. commercial revenue. CrowdStrike outperformed, and Jim Cramer bought it for his fund, arguing that cybersecurity is harder to replicate with agents than workflow automation. Apple dodged the selloff entirely. Hardware companies, data infrastructure providers, and security firms were spared. The market was discriminating. It wasn't selling "tech." It was selling seats.

The Per-Seat Model Was Already Cracking

The selloff didn't create a problem. It ripped the bandage off one that had been festering since at least 2022. Per-seat pricing built modern SaaS, but the cracks were already spreading before Anthropic shipped a single plugin. AI agents didn't create the cracks. They gave the market a reason to stop ignoring them.

About 40% of SaaS companies still price by the seat, according to industry surveys. The model is simple. Each human user represents a unit of revenue. More employees using the software means more revenue. Growth tracks headcount expansion. Wall Street ate it up. Predictable. Recurring. Sticky as hell. Switching costs kept churn low.

But the model contains a structural assumption that nobody questioned until recently. It assumes humans operate the software.

Jason Lemkin, the SaaStr founder who has tracked SaaS economics for over a decade, offered a contrarian read. "This isn't the death of SaaS," he wrote on February 4. "It's the end of easy SaaS." His argument was that the crash represented the market pricing in growth deceleration that had been building since 2021, and that AI gave Wall Street a narrative to pin the selloff on. Growth was already slowing. Price increases, not new customer acquisition, were driving revenue. At Salesforce, 72% of 2025 revenue growth came from raising prices on existing customers rather than adding new ones.

Lemkin has a point about timing. But the structural threat is real even if the selloff's intensity was sentiment-driven. Consider the arithmetic. Run the math. Ten AI agents handling what 100 sales reps used to do. Ninety percent fewer seats. Revenue craters unless the vendor rewrites the entire pricing playbook. One Base44 customer killed a $350,000-per-year Salesforce contract after rebuilding the whole thing with AI tools. Not a whiteboard exercise. A check that stopped clearing.

SaaS cost inflation had been running at 11.4% in 2025, four times faster than G7 inflation. Zylo's annual index found enterprises spending an average of $49 million on SaaS annually, with $21 million wasted on unused licenses. The seat model was already extractive. Buyers were looking for an exit. Agents gave them one.

Revenue multiples tell the longer story. SaaS EV-to-revenue ratios peaked at 18 to 19 times during the pandemic. By December 2025, the median had fallen to 5.1 times. The percentage of software companies beating revenue estimates dropped to 67%, down from a historical norm above 75%. Wall Street's premium for predictable per-seat recurring revenue was evaporating before Anthropic shipped anything.

What Cowork Actually Replaces (and What It Doesn't)

The market's reaction was broad, but the technical reality is more specific. Cowork and its plugins occupy the workflow layer of enterprise software. They handle tasks that are repetitive, document-heavy, and tolerance-flexible. Contract review. Lead management. Expense processing. Project coordination. Tier-1 customer support resolution. This is the layer where "close enough" is good enough, and where per-seat SaaS commands its highest markups.

Bain & Company's 2025 technology report drew a useful distinction between what it called "probabilistic SaaS" and "deterministic SaaS." Probabilistic software handles tasks where approximate answers are acceptable. Summarizing a legal brief, drafting a marketing email, triaging a support ticket. AI agents are very good at these. Deterministic software handles tasks where 100% precision is required. Closing a financial ledger, randomizing a clinical trial, processing a payment. AI agents are not good enough for these, and may never be, because the failure mode isn't a bad summary. It's a compliance violation.

The selloff hit probabilistic categories hardest. Monday.com, Atlassian, HubSpot, LegalZoom. Deterministic categories were largely spared. Procore's project cost accounting, Medidata's clinical-trial systems, payment processors. CrowdStrike actually rose during the worst of it. Security software is not just deterministic. It's adversarial. The attacker adapts. That makes it harder to replace with agents trained on historical patterns.

Opus 4.6's benchmarks explain why the market reacted the way it did. The model scored 65.4% on Terminal-Bench 2.0, the highest result for any agentic coding benchmark. On GDPval-AA, a finance and legal task evaluation, it posted an Elo rating of 1606, a 144-point lead over GPT-5.2. These are not demonstration numbers. They represent production-grade performance on exactly the tasks that per-seat SaaS charges premium rates for.

But the technical reality includes limits that the market selloff didn't price in. Agents still struggle with multi-step reasoning across ambiguous contexts. They make things up. Somebody still has to watch them when the stakes are real. And those 11 Cowork plugins? Open-source, which in enterprise terms means raw material, not a product you'd bet your compliance on. No SOC 2 compliance. No role-based access controls. No audit trails. Enterprises buying Thomson Reuters or ServiceNow are paying for reliability guarantees, not just functionality. That gap won't close overnight.

There's a subtler point the market missed entirely. Workflow SaaS companies don't just provide functionality. They provide institutional memory. A company's Salesforce instance contains years of customer relationship data, custom objects, automated workflows, and integrations with dozens of other systems. An AI agent can draft an email or triage a support ticket on day one. It cannot replicate a decade of accumulated process logic and data relationships. The switching cost isn't the software. It's the organizational knowledge encoded in how the software is configured.

This is why Janus Henderson's Cofsky argued that "the narrative ignores the breadth of these legacy entities' business models." SaaS companies that have become deeply embedded in enterprise workflows have a moat that is not visible in a product demo. The question is whether that moat is wide enough to justify current valuations, or merely wide enough to slow the decline.

The New Math: Four Ways to Price Software Without Seats

If per-seat pricing is breaking, what replaces it? The answer is already shipping. A handful of SaaS companies have already started crossing over. Early numbers look workable. The process, though, is ugly.

Kyle Poyar left OpenView to run Growth Unhinged, and he counted more than 1,800 pricing changes across the top 500 SaaS and AI companies last year alone. "2025 will be remembered as the year when seemingly everybody lost confidence in their pricing," he wrote in January 2026.

Poyar and Manny Medina analyzed over 60 companies and identified four emerging models.

Per-agent FTE replacement. Companies like 11x and Harvey charge a fixed monthly fee per deployed agent, positioning the agent as a cheaper alternative to hiring a human. The strategic insight is that this targets the headcount budget, which is typically 10 times larger than the technology budget. The software isn't competing with other software. It's competing with salaries.

Per-agent action. Salesforce's new Flex Credits system charges $0.10 per action. Organizations buy credits in blocks of 100,000 for $500. The Flex Agreement lets customers convert existing user licenses into credits or vice versa, creating a direct bridge from the old model to the new one. It's the clearest admission from a major SaaS company that seats are a transitional artifact. Agentforce pulled in 18,500 deals in year one. Of those, 9,500 were paid, pushing the product to $540 million in ARR. That's 330% growth year over year.

Per-agent workflow. Companies like Rox charge for completed task sequences rather than individual actions. This sits between action-based and outcome-based pricing, capturing more value than raw consumption while avoiding the definitional challenges of "outcomes."

Per-agent outcome. Intercom charges $0.99 per resolution. Only when the customer's issue is actually resolved. Zendesk charges $1.50 to $2.00 per automated resolution. Intercom's AI agent, Fin, now resolves 56% of conversations autonomously, up from roughly 25% a year ago. Intercom's revenue quadrupled in twelve months on that bet.

Tomasz Tunguz wrote something in July 2024 that nobody paid much attention to at the time. It reads like prophecy now. "What does a software seat mean when a human is no longer operating the software?"

Lemkin's own company offers the most concrete case study of what the transition looks like from the buyer's side. SaaStr deployed more than 20 AI agents in production, replacing what had been a 10-person go-to-market team with 1.2 humans. The agents sent over 60,000 hyper-personalized emails, booked 130-plus meetings automatically, and generated 15% of the company's London event revenue without human involvement. Open rates on ghosted leads hit 70%. SaaStr maintained eight-figure revenue with single-digit headcount. "We're already downgrading our seat counts at vendors," Lemkin wrote, "because we have 12-plus AI agents in production. We just have fewer humans and more AI agents."

That last sentence is the one SaaS CFOs should tape to their monitors. Fewer humans, more agents, fewer seats. The revenue unit shrinks with every agent deployed. And the agent deployment curve is steepening. Gartner thinks 40% of enterprise applications will ship with task-specific AI agents built in by December. Last year, fewer than 5% did. An eightfold jump in one calendar year, if the forecast holds.

Copilot is the cautionary tale here. Microsoft priced it at $30 per user per month when it went live in late 2023, and the enterprise market basically yawned. Now, two years in, only 3.3% of the commercial base has signed up. Fifteen million paid seats out of a 450 million commercial base. Microsoft has since cut the price to $18 for small businesses and added Azure-metered agent access on top. The hybrid approach, seats plus consumption, is an implicit admission that pure per-seat AI pricing failed to reach scale.

The broader pattern is clear. AI-native companies operate at 5.7 times the revenue efficiency of traditional SaaS. Revenue per head tells the story. AI-native firms average $3.48 million per employee. Traditional SaaS? $610,000. Gamma crossed $100 million ARR with 50 people on staff, which would have been science fiction five years ago. And the barriers to entry have evaporated. Building a competing SaaS product used to cost $50,000 to $500,000 and take six to eighteen months. Now it's $500 to $20,000 and two to twelve weeks. The economics of building a seat-based competitor used to require venture scale. Now it requires a credit card.

The buying side noticed. A CIO.com analysis from late last year found enterprise tech leaders already renegotiating SaaS contracts, stuffing in usage ceilings, agent-interaction limits, and delay clauses that didn't exist twelve months ago. FinOps teams are deploying monitoring tools to track AI agent activity the way they once tracked cloud compute costs. One emerging concern is particularly telling. Some CIOs worry that usage-metered AI pricing creates a new attack surface. If agents are priced per action, a denial-of-service style attack that triggers millions of agent actions could inflate a company's software bill the way a DDoS attack can inflate hosting costs. The security implications of consumption-based AI pricing have not been worked out. Nobody is waiting for them to be.

Collateral Damage: India's $300 Billion Industry and Wall Street's Private Credit Exposure

The selling didn't stay in Silicon Valley. It spread to Bangalore and to Wall Street's private credit desks, revealing how deeply the per-seat model is wired into global economic structures that have nothing to do with software.

India's IT services industry is worth $300 billion and accounts for 10% of the country's GDP. Five million people work in it directly. The entire business model rests on two foundations. Labor arbitrage, where Indian engineers cost 60 to 80 percent less than their Western counterparts, and billable hours, where firms charge clients per person per month on long-term engagements.

AI agents threaten both pillars simultaneously. "The timelines of engagements will massively shrink further, impacting billing," said Yugal Joshi of the Everest Group. Ishan Talathi, co-founder of CloudPe, put it more directly. "The Indian IT model is built on man-day billing. We charge for bodies on projects. That model is now facing existential pressure."

Up to half of traditional outsourcing work is at risk, according to Market Research Future analysis. TCS pegs its AI services revenue at $1.8 billion, roughly 5% of quarterly consolidated revenue. The top five Indian IT companies have trained more than 250,000 employees in AI. But training employees in AI and replacing billable hours with AI are opposite strategies. One adds capability. The other destroys the revenue unit.

Accenture's stock fell nearly 10% on February 4 despite posting Q1 revenue of $18.74 billion, beating estimates by more than $200 million. The company has been acquiring AI firms aggressively. Faculty, Profitmind, others. Bookings of $20.9 billion suggest demand remains strong. But the consulting model faces the same arithmetic as per-seat SaaS. If AI can draft, summarize, and analyze faster than a team of junior consultants, clients require less external help on the routine work that drives billable hours. The high-value strategy work survives. The implementation hours do not.

The private credit angle is less visible but potentially more dangerous. UBS analysts estimate that 25 to 35 percent of the private credit market is exposed to AI disruption risk. Software accounts for roughly 20% of outstanding loans at major private direct lenders. Most of those loans were written in 2022 and 2023, when valuations were still bloated and everybody still believed the per-seat growth story.

Default rates in private credit are expected to climb roughly two points to around 6% this year. And here's where it gets recursive. Software company valuations drop, which shrinks collateral values. Collateral shrinks, covenants trip, margin calls go out. Lenders force sales, which push prices down further. More covenants trip. The mechanism is familiar from other credit cycles. The novelty is that the underlying collateral is recurring software revenue, which was supposed to be the safest asset class in technology.

Advertising agencies were hit from a different direction. Barclays surveyed buy-side investors and found that agencies were viewed as the "most exposed" segment of European media to AI disruption. Omnicom fell 11% in a single session. WPP lost 10% and then fell another 3.7% the next day. Publicis dropped 9% despite posting record operating margins of 18.2%. Havas lost 7.5%.

The agency selloff has a concrete precedent that makes it harder to dismiss. Klarna had already ripped out Salesforce and Workday, axed 1,500 employees, and slashed marketing agency spending by 25%, pocketing $4 million. That's one mid-sized fintech client executing the playbook. The sell-side is pricing in what happens when hundreds of enterprise clients make the same calculation.

The compounding effect matters here. When a single company like Klarna pulls out of SaaS contracts, dials back consulting engagements, and fires its agencies all at once, the damage bleeds across three different sectors simultaneously. The SaaSpocalypse was not really about software alone. It was about every business model that charges by the unit of human labor, whether that unit is called a seat, a billable hour, or a campaign retainer.

The Bull Case: BofA's Logical Contradiction and the Adobe Playbook

Goldman Sachs' Matthew Martino cautioned clients in a note on February 5 that "the recent selloff reflects a rapid shift in investor sentiment rather than a sudden deterioration in fundamentals." His colleague Ben Snider added a more careful qualifier. "The uncertainty around the eventual impact of AI means near-term earnings results will be important signals of business resilience, but in many cases insufficient to disprove the long-term downside risk." In other words, good quarters won't save you if the model is structurally impaired. That's the bearish read even from a bank telling clients not to sell.

Bank of America's Vivek Arya published the sharpest rebuttal to the selloff on February 4. His argument was not that SaaS would survive. It was that the market was pricing in two mutually exclusive scenarios at the same time.

Investors, Arya wrote, were simultaneously pricing in AI capital expenditure deteriorating to the point of weak returns, and AI adoption being so powerful that it makes legacy software business models obsolete. Both cannot be true. If infrastructure spending collapses because AI doesn't deliver returns, the technology cannot be powerful enough to destroy SaaS. If AI is powerful enough to destroy SaaS, infrastructure spending cannot collapse. Pick one.

Janus Henderson's Jonathan Cofsky added context on valuation. Application software stocks were trading at roughly 20 times 2027 earnings by early February, well below historical averages and below the broader market multiple. "AI versus SaaS is not an either/or proposition," Cofsky wrote. "There is room for both to thrive."

The strongest bull case is historical. Adobe's transition from perpetual licenses to Creative Cloud subscriptions between 2012 and 2015 is the closest parallel to what SaaS companies face today. Before the transition, Adobe generated $4.4 billion in annual revenue. Its stock traded below $30. When it moved to subscriptions, revenue went flat. Net income collapsed from $833 million to $268 million. The stock dropped 12%. Analysts questioned whether the company would survive.

Three years later, revenue exceeded the pre-transition peak. By 2020 the company was pulling $12.9 billion a year, three times what it made before the switch. Over 90% recurring. Thirty-seven million subscribers on the rolls. If you bought Adobe in 2012 and held until 2023, your return was 1,200%. You'd have made thirteen times your money, assuming you didn't panic-sell during the three years when every analyst on the Street was asking whether the company would survive.

That's the part bulls keep coming back to. The companies that ripped the Band-Aid off early and committed to subscriptions came out structurally better. The hedgers, the ones who kept perpetual licenses around to protect near-term revenue, watched their market share drain away and never got it back.

But the Adobe analogy only stretches so far. Adobe controlled its transition. It was moving from one pricing model it owned to another pricing model it owned. The product stayed the same. Photoshop was still Photoshop. Only the billing mechanism changed.

SaaS companies facing AI agents don't have that luxury. They're not just changing how they charge. They're defending against external competition that operates on fundamentally different economics. Standing up a competing SaaS product went from a six-figure project to a weekend with a credit card. Development timelines collapsed from over a year to a few weeks. An AI-native startup with $3.48 million in revenue per employee isn't just cheaper. That's not a cheaper competitor. That's a different animal entirely. Nobody was cloning Photoshop in two months for five grand when Adobe made its move.

But the pattern holds in one important respect. Who survives the per-seat-to-per-outcome transition? Companies sitting on proprietary data, companies in regulated verticals where compliance isn't optional, and companies that have already started rewiring their pricing. Salesforce, with Agentforce at $540 million ARR and a Flex Credits bridge mechanism, is further along than most. ServiceNow is repositioning as an "enterprise AI operating system" and targeting $1 billion in AI revenue by 2026. Both companies fell during the selloff. Both will likely recover.

The companies that won't recover are pure workflow tools with no data moat, no platform play, and no pricing flexibility. Monday.com's 50% decline is telling. The company beat earnings. Wall Street doesn't care about a good quarter when the model is cracked. One AI agent coordinates projects, assigns tasks, and tracks deadlines for a flat $200 a month. Stack that against $19 per seat across a company with hundreds of employees. The arithmetic is brutal, and it eats the revenue model alive. Beating the quarter doesn't fix a broken revenue unit.

The Third Act of Software

A useful frame for what comes next is to stop thinking about SaaS as a product category and start thinking about it as a pricing era. SaaS has moved through two structural transitions in 25 years. The first shift, from on-premise perpetual licenses to cloud subscriptions, ground on for about a decade. Oracle and SAP made it through, though the transition wrecked their cash flows for a while. Most of the mid-tier vendors never made it out alive. Then came act two, roughly 2015 to 2025. The subscription economy. Per-seat pricing. Revenue multiples peaking at 18 to 19 times. The era of predictable, recurring, growing revenue streams that Wall Street loved.

The third act started on February 3, 2026, when the market decided the second act was over.

IDC's forecast is blunt: pure seat-based pricing hits obsolescence by 2028, and 70% of vendors will have to rethink what they're actually selling. Gartner stretches the window to 2030 but arrives at a similar place, projecting 40% of enterprise SaaS spend shifting to usage, agent, or outcome models. And Deloitte's 2026 tech predictions? Seventy-five percent of companies say they plan to invest in agentic AI. Half of those organizations intend to pour more than 50% of their digital transformation budgets into AI automation.

These are not speculative estimates. They're trend extrapolations from transitions already underway. Salesforce has a working outcome-based product with $540 million in ARR. Intercom quadrupled revenue in a year on per-resolution pricing. Zendesk launched hybrid seat-plus-outcome plans. Microsoft, after two years of struggling with per-seat AI pricing, has added consumption-based layers. The direction is set. The question is speed.

The $285 billion selloff was not the market panicking about artificial intelligence. It was the market doing arithmetic. When agents do the work, seats lose their value. And when seats lose their value, so do the revenue multiples that justified software stock prices for a decade.

SaaS does not die. The software itself isn't going anywhere. Someone has to write the code, patch the vulnerabilities, satisfy the auditors. Regulated industries still need audit trails. Security still requires adversarial thinking that agents can't replicate. Data infrastructure sits underneath everything and grows more valuable as agent adoption creates new categories of risk, complexity, and governance demand. The picks-and-shovels plays will thrive. Compliance platforms. Identity and access management. Those businesses grow regardless.

But the transition will not be orderly. AlixPartners expects software M&A to surge 30 to 40 percent year-over-year in 2026, with deal values reaching $600 billion. Some of that will be consolidation among survivors. Some will be fire sales. The companies stuck in no-man's-land are the ones that won't make it. Too small to become platforms. Too hooked on per-seat revenue to stomach a pricing overhaul. Too generic to charge a premium. They'll get bought at fire-sale multiples, or they'll just go dark. Venture-backed SaaS companies that raised at 2021 valuations and still aren't profitable feel this most acutely. Their investors are already underwater. Their growth is decelerating. And now their pricing model is broken.

What dies is the assumption that a human sitting in a chair clicking buttons is the right unit of software revenue. That assumption built a $390 billion industry. AI agents just broke it. What comes next will be messier, more volatile, and ultimately more valuable. But only for the companies that stop charging for chairs.

What to Watch

Salesforce earnings (February 26): Agentforce deal count and Flex Credits adoption rate. If seat-to-credit conversions accelerate beyond the current 18,500 deal pace, the pricing transition is entering a mainstream phase. The tell will be in average deal size commentary. If execs start quoting per-action metrics instead of per-seat, the transition is real.

Microsoft Copilot penetration by mid-2026: Still stuck at 3.3% of the commercial base after two years of pushing and a price cut from $30 down to $18. If that number hasn't budged by summer, the per-seat AI add-on model is finished, and Microsoft's Azure-metered agent play becomes the only AI monetization path that works.

Indian IT Q1 guidance (April-May): TCS, Infosys, and Wipro forward guidance will reveal whether clients are actively renegotiating billable-hour contracts or simply pausing new engagements while they figure out agent deployment. The difference matters. Pauses are temporary. Renegotiations are structural.

Private credit software loan performance through Q2: Watch default rates in software-backed loan portfolios. If they exceed the projected 6%, the contagion from SaaS repricing extends into financial markets in ways that amplify the original selloff.

Monday.com and Atlassian product roadmaps: Both need to ship outcome-based or consumption-based pricing tiers before summer, or investor confidence won't recover. If either company is still running pure per-seat by year-end, expect the stock to keep bleeding.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai