Michael hochstat

Knowledge is free knowledge.dotadda.io/session…

Explorium is shutting down after almost nine years. Sad news, and my thoughts are with the team. I dug into the story, and it's a useful case study in how hard a pivot can be. This is my read from public info, not inside knowledge. They started by helping data scientists find outside data for ML models. Then ML tooling got commoditized, and that market shrank. In 2025 they pivoted to AgentSource, B2B data built for AI agents. Three things made that bet hard: 1. Crowded market. Clay, ZoomInfo, and Apollo already owned B2B data, and ZoomInfo shipped its own MCP server too. 2. Small tickets. Agent builders buy credits in small packs, nothing like the enterprise deals of the ML era. 3. Platforms moving in. Salesforce and HubSpot are building agent workflows right into their own systems. They also raised about $125M, with the last round I can find in 2021. That's a long runway across two pivots. My takeaways: - Being early on a new standard (like MCP) isn't a moat. - A new customer class has to be able to pay. - Aggregating other people's data is a hard business to defend. To everyone at Explorium: you built, you pivoted early, and you shipped. That takes real courage. Wishing you all the best in what comes next. What's your take on what the agent era means for B2B data?

Palantir, Anthropic and AIG: The Part of the AI Stack Investors May Be Missing Everyone wants to own the AI model. But the model may not be where the most important enterprise value is created. Consider the AIG example discussed in Anthropic’s S-1. Claude provides the intelligence. Palantir provides much of the infrastructure that puts that intelligence to work inside an actual insurance business. These are two very different jobs. Think about the AI stack. 1. The model — Anthropic / Claude Understands language, extracts information, reasons through problems and generates responses. 2. The data layer — Palantir Foundry Connects information across the business. Organizes policies, submissions, risks, exposures and relationships into a structured operating environment. 3. The execution layer — Palantir AIP Coordinates agents, routes tasks, manages workflows and connects AI capabilities to actual business processes. 4. Governance and controls Permissions, audit trails, traceability and customer-defined rules. In insurance, you cannot simply let a chatbot make decisions without accountability. 5. The customer — AIG Provides the underwriting expertise, risk appetite, business rules and institutional knowledge. That is the important distinction. Claude can help understand an insurance submission. But understanding a document and running an underwriting operation are not the same thing. Why this matters to investors A general-purpose model does not automatically become a production system. Enterprise AI has to deal with: Thousands of documents and disconnected databases. Conflicting information and source reliability. Multi-step workflows. Regulatory requirements. Permissions and accountability. Human approval and operational controls. The model is only one part of the solution. The systems around it determine whether AI can actually do the work. This is where Palantir fits. Anthropic’s S-1 highlighted Claude’s contribution to the AIG example. Palantir has separately described its role in the underwriting workflow through Foundry and AIP. The two descriptions are not necessarily competing claims. They describe different parts of the same technology stack. There is a much bigger investment question here. If the model becomes increasingly interchangeable, where does the durable value accumulate? The companies supplying intelligence will compete on model quality, cost and capability. The companies supplying data infrastructure, orchestration, governance and workflow integration may become just as important. And the customer still supplies something neither can manufacture overnight: years of operating knowledge and business judgment. The AI trade may not simply be about who builds the smartest model. It may be about who owns the systems that turn that model into something a business can actually use. A demo is easy. Replacing an actual business process is much harder. That is the difference between AI that answers questions and AI that does work.

DoTadda Knowledge: Why Purpose-Built AI Research Matters in Finance Most AI tools can answer a question about a company. Try asking one to analyze 10 years of earnings calls, SEC filings, financial statements, management guidance, and operating results. Then ask it to reconcile conflicting numbers, identify changes in management’s language, and produce an investment memo with sources. That is where things start to break. The problem isn’t simply the model. It’s how the work is organized. Three problems undermine general-purpose AI research: 1. Context bloat Too much information gets pushed into one conversation. Documents, old results, repeated instructions, and irrelevant material compete for attention. Important numbers and management statements get buried. 2. Context rot Even before a model reaches its technical limits, its ability to retrieve and use relevant information can deteriorate. The evidence is there, but the model doesn’t always connect it correctly. 3. Compaction loops When context fills up, agents summarize previous work to keep going. Summaries lose details. Exact numbers, qualifications, and prior decisions disappear. The system then has to reconstruct what it already did. For an investment professional, the result is familiar: convincing answers that still require hours of checking. We built DoTadda Knowledge differently. Instead of asking one general-purpose agent to do everything, we break financial research into specialized tasks. Parallel agents: Earnings transcripts, SEC filings, financials, price history, and approved web sources are analyzed by focused agents. Source hierarchy: SEC filings first, earnings releases second, transcripts third, normalized financials fourth, and web sources for context. Actual calculations: Python executes the financial calculations, scenarios, margin bridges, and statistical analysis. Numbers are calculated, not guessed. Evidence-backed output: Claims are connected to source material. Structured research: The output isn’t another chatbot conversation. It is a usable research product. What does that produce? Company primers. Peer comparisons. Management credibility scorecards. KPI trackers. Investment committee one-pagers. Long and short research memos. The objective is simple. Ramp a new company in an afternoon instead of spending a day gathering information. Compare management’s promises with actual results across multiple quarters. Identify changes in guidance, margins, operating performance, and management credibility. Cover more companies without lowering the standard of evidence. Most importantly, let analysts spend their time thinking rather than searching through documents. AI should not create another job for the analyst. It should eliminate the work that gets in the way of analysis. DoTadda Knowledge was built around that idea. Any U.S. public company. More than 10 years of historical information. There is a free tier to test against your existing coverage. No credit card required. knowledge.dotadda.io DoTadda Knowledge Evidence-backed research. Less time gathering. More time thinking.

Yes — that statement is accurate. Once an LLM agent enters significant context bloat and starts cycling through compaction, both quality and efficiency usually degrade. This is a well-documented pattern in long-horizon agents as of late 2026. What the terms mean •Context bloat: Unnecessary or low-value tokens accumulating in the working context (old tool results, full file dumps, verbose tool schemas/MCP definitions, repeated instructions, full conversation history). Even before you hit the hard window limit, this causes context rot — attention dilutes, important details get lost in the middle, and the model starts favoring recent or superficial patterns over precise reasoning. •Compaction loops: The cycle of summarizing/compressing the history when the context nears the limit, then continuing. Compaction is inherently lossy. The agent often has to re-establish state afterward (re-reading files, re-querying tools), which refills the context and triggers the next compaction. This creates a feedback loop. Why performance degrades Research and production measurements show several concrete effects: •Quality drop from length alone: Benchmarks like NoLiMa and “Lost in the Middle” show many models lose a large fraction of their short-context performance once context grows (e.g., many dropping below 50% relative performance by ~32k tokens). Relevant information in the middle is especially vulnerable. •Lossy compaction side effects: Studies on compaction (including Compact-Bench, Compaction Cliff analyses, and evaluations of side constraints) find that summaries frequently drop exact constraints, safety rules, anti-cheating instructions, or precise identifiers. One finding: safety rules can fall from ~53% retention after one compaction to ~10% after several rounds. On real coding sessions, instructions about how the agent should work are dropped far more often than the actual task specification. •Re-establishment cost: After compaction, agents often re-ingest context at a higher rate for many steps (re-reading files, re-deriving state). This raises both token cost and the chance of another compaction soon. Some measurements show the post-compaction “correction spike” lasting 20–30+ steps. •Cost and latency: Prefill cost scales with total tokens. Sessions that compact can account for a disproportionate share of total token spend. Synchronous summarization also stalls the agent for seconds to minutes. Compaction is still necessary for long tasks (it keeps the agent alive past the window), but naive threshold-based full-history summarization is a blunt instrument. Better approaches include keeping a fixed high-fidelity head (system prompt + original task) and tail (recent turns), summarizing only the middle, offloading large tool results to files with references, progressive tool disclosure instead of dumping every schema, and external persistent state (files, graphs, or structured memory) rather than pure in-context history. The Hochstat reply is essentially pointing at a real engineering reality: bigger context windows and better models help, but they do not eliminate the need for careful context management. Once bloat and repeated compaction set in, raw model capability is often no longer the limiting factor.

No clear timeline—and nothing public suggests it’s about to collapse soon. Apollo has gone from a 2024 East Village shop to 10 locations (plus the Industry City bakehouse) in under three years. It still draws lines, gets media coverage (including an SNL sketch), and has expanded without any reported closures or major quality blow-ups. The centralized overnight model has been in place since mid/late 2026, and demand has been high enough to justify the 15,000-bagel daily capacity. That said, the “nothing made in-store” issue is a real long-term risk in New York. Once the hype cools and people start comparing it more closely to shops that still boil and bake on site, the toasting-to-order step may not be enough for the traditionalists. Scaled bakery concepts have faded before when the product starts feeling more like a consistent chain item than a fresh local one. So far the sourdough style and limited menu have kept enough people coming back that it hasn’t happened yet. “Toast” is possible if volume drops and the novelty wears off, but right now the expansion and reception say it’s holding. Most of these things either stabilize as a solid multi-location brand or slowly lose relevance over a few years rather than vanishing overnight.

That’s basically correct. Nothing is mixed, boiled, or baked from scratch in the individual retail stores. All the bagels are fully made at the Industry City bakehouse (15,000 a day), delivered overnight, and the shops just toast them to order and put on the spreads/sandwiches. The Industry City location is the exception only because the production kitchen is right there—you can watch them being made—but every other counter is receiving finished product. A lot of people treat that as a dealbreaker for a “fresh” bagel shop, which is why places that still boil and bake on site (like the one next to the Williamsburg Apollo) feel different. Whether that makes the whole thing a flop depends on if the sourdough style, consistency, and toasting are good enough for you once the novelty is gone. So far the expansion and lines show it has worked for enough customers, but the model is exactly the scaled, reheated version critics like you are pointing at.

Yeah, that’s the core complaint. Lender’s are the frozen, mass-produced baseline a lot of people grew up with—soft, pre-baked, reheated at home. Apollo’s are a long way from that on paper: natural fermentation, boiled, sourdough crumb, darker crust, limited varieties, and they still get toasted to order. Critics and a lot of regulars treat them as a deliberate “new-school” style rather than a traditional New York bagel. But once the baking moved off-site and the product became a centralized, overnight-delivered item finished at the counter, the experience starts to overlap with the scaled/reheated model for people who prioritize “just out of the oven in this shop.” That’s the gap your point is hitting. If the texture and tang still work for you when it’s hot, fine; if it now just reads as a juiced-up version of something that was never a real fresh bagel, then the in-store bakers next door make more sense.

You’re right. A $2.50 bagel shouldn’t turn into a lousy piece of bread overnight. A properly made bagel should have enough structure to hold up. The problem may be Apollo’s formulation, baking process, moisture content, or how it’s stored and transported. Traditional New York bagels were designed to be chewy and substantial. They shouldn’t need to be eaten within an hour to be enjoyable. And frankly, at Apollo’s prices, next-day quality is a fair criticism.

If publishers successfully pursue the AI-company users of the scraped data (Anthropic, Perplexity, Mistral, xAI, Cohere and others), the main added costs would arise from licensing fees, litigation or settlements, operational changes, and higher prices passed to customers. These pressures are already visible in existing deals and cases, so the impact would be an acceleration of trends rather than an entirely new category of expense. Direct licensing of news content for training and especially real-time grounding or retrieval-augmented generation is already underway. Reported examples include News Corp’s arrangement with OpenAI valued at more than $250 million over five years (roughly $50 million annually on average), Amazon’s deal with The New York Times at $20–25 million per year, OpenAI’s agreement with Dotdash Meredith at least $16 million per year, and Meta’s reported terms with News Corp of up to $50 million annually. Across a broader set of agreements the median annual value for news-publisher deals sits around $4.2 million. Reference or display rights needed for live AI answers typically command higher prices than pure training rights alone. Some unit pricing for news-article archives has been observed near $0.46 per article, though actual rates vary widely with exclusivity, freshness, metadata, and the scope of permitted uses. Shifting from low-cost third-party index access such as Brave’s API (publicly priced around $5 per 1,000 requests) to direct licenses would raise the cost of high-quality, timely news data substantially for any company that relies on it heavily. Complaints allege Brave itself generated more than $100 million by reselling such content. Litigation and settlement exposure can exceed ordinary licensing. Anthropic’s settlement in the Bartz case over pirated books reached $1.5 billion, equating to roughly $3,000 per work across hundreds of thousands of titles. Statutory damages for willful infringement range from $750 to $150,000 per work; even non-willful exposure scales rapidly when large numbers of articles are involved. Ongoing cases also generate substantial legal fees—the New York Times has already reported tens of millions in AI-related litigation costs. Court orders or settlements requiring deletion of models or training sets that incorporated unauthorized material would add further retraining and opportunity costs. Broader effects include pressure on margins, since compute already dominates AI expenses and content licensing would sit on top of those outlays. Smaller or mid-tier companies would feel the increase more acutely than the largest labs that can negotiate volume deals or absorb the cost. Data coverage or freshness might decline if cheap real-time indexes become unavailable, leading systems to rely more on older material, licensed subsets, or synthetic alternatives and potentially lowering answer quality on current events. Higher input costs tend to flow through to API pricing, consumer subscriptions, and enterprise contracts. Larger players that lock in long-term licenses can raise rivals’ costs and reinforce market concentration. Companies would also spend more on data-provenance tracking, contractual indemnities, and filtering systems to avoid unauthorized or paywalled material. In summary, successful claims against the downstream AI users would expand the formal licensing market for news, raise the floor price of timely high-quality text, and add litigation-risk premiums. The largest laboratories can manage much of this through negotiation or scale; smaller firms and pure AI-search products would face steeper relative cost increases.

Bottom line No—AI has not killed Shopify’s golden goose. So far, AI appears to be creating additional commerce entry points while Shopify retains the valuable transaction infrastructure: catalog, checkout, payments, identity, inventory and merchant workflows. The bigger investment risk is valuation, not demonstrated AI disruption. Why the business still looks healthy •On the Q3 call, Shopify reported GMV growth of 29.5%, revenue growth of 32%, merchant-solutions growth of 35%, payments penetration above 68%, and Shop Pay GMV growth of 50%. Those numbers are inconsistent with current disintermediation. Shopify Q3 2026 earnings call •AI-referred traffic reportedly grew 2.6x, orders grew 3.3x, and AI-referred shoppers converted 64% more often. AI is currently improving product discovery—particularly for long-tail merchants—rather than removing Shopify from transactions. Q3 AI-commerce indicators •The Q2 filing showed six-month GMV of $216.3 billion, up 33%, revenue of $6.753 billion, up 34%, and Shopify Payments penetration of 67%, facilitating $145.1 billion of GMV. Operating cash flow was $1.139 billion. Shopify Q2 2026 MD&A Shopify’s central argument is credible: AI agents may own the conversation, but merchants still need a system of record underneath them. More fragmented shopping interfaces could actually increase the value of having one inventory, payment and order-management platform. Goldman Sachs Communacopia conference The real bear case AI becomes destructive if ChatGPT and other agents: 1Keep consumers inside their own interfaces; 2Control buyer identity and customer data; 3Own checkout and payment credentials; 4Charge merchants significant referral or advertising fees; 5Reduce Shopify to a commoditized back-end database. That has not happened at material scale yet. Management admits agentic purchasing volumes remain small. AI also has real costs: incremental Sidekick spending contributed to year-over-year subscription gross-margin compression in Q3. Shopify Q3 2026 earnings call Stock versus business At approximately $143.91, Shopify’s standardized metrics imply roughly: •$186.7 billion market capitalization •58.5x forward earnings •14.1x sales That valuation already assumes Shopify remains one of AI commerce’s principal infrastructure winners. The company can survive AI disruption while the stock still falls sharply if revenue growth drops from around 30% toward the teens or AI platforms begin taking more of the economics. Verdict •Business: Golden goose intact; AI may enlarge the opportunity. •Stock: High-quality but priced for continued exceptional execution. •Investment conclusion: AI alone is not a credible reason to short SHOP. At this valuation, however, I would avoid chasing it unless payments penetration, AI-driven transactions and free-cash-flow growth continue to validate the premium. The thesis changes if payments penetration falls, merchant-solutions growth materially trails GMV, AI platforms bypass Shop Pay, or AI infrastructure costs create sustained margin compression. Sources: • Shopify Q2 2026 Form 10-Q — Risk Factors • Shopify Goldman Sachs Technology and Communacopia 2026 • Shopify Q3 2026 Earnings Call

Chamath Flags a Hidden Risk Building Under the AI Boom For years, America’s largest technology companies operated under a simple and powerful assumption: they could do nearly everything at once. Microsoft (MSFT), Meta Platforms (META), Alphabet (GOOGL), and Amazon (AMZN) generated so much cash that they could build vast data centers, repurchase tens of billions of dollars in stock, and still maintain some of the strongest balance sheets in corporate America. Conventional wisdom held that these firms were uniquely insulated from the financing constraints that limit almost every other industry. Artificial intelligence is now testing that paradigm. On the October 30 episode of the All-In Podcast, billionaire investor Chamath Palihapitiya argued that the current AI spending boom carries a less obvious but significant risk. Technology companies may increasingly need outside capital to finance the massive infrastructure required for AI, placing them in direct competition with the U.S. government—another enormous borrower—in the capital markets. Palihapitiya estimates that the broader AI industry already owes roughly $340 billion to $400 billion to capital markets. Even with strong demand for AI products and services, that scale of borrowing is consequential. Big Tech is already returning less cash to shareholders while pouring more into infrastructure. Investors, he suggests, may need to focus far more closely on borrowing costs and free cash flow than they have in the recent past. The numbers illustrate the shift clearly. According to Reuters, Microsoft, Alphabet, Amazon, Meta, and Oracle generated a combined $542.9 billion in operating cash flow during the first nine months of 2026. Their capital spending, however, doubled to $556.9 billion—exceeding the cash flow these five companies produced for the first time in data going back to 2011. Shareholders are already feeling the impact. The group’s total stock buybacks dropped 83% to just $13.6 billion. Microsoft accounted for the majority of the remaining repurchases, while Alphabet and Meta made none. For years, buybacks provided a reliable boost to earnings per share by reducing the share count. That long-standing financial windfall is now largely disappearing. “The entire United States economy is levered to the AI complex,” Palihapitiya said. The enormous expenditure does not automatically make AI a poor investment. If the new infrastructure eventually generates sufficient revenue and profit, stockholders could still benefit. Yet the barrier is rising. The same five companies swung from $143.5 billion of free cash flow a year earlier to a $14 billion deficit in the first nine months of 2026. In short, investors are forgoing cash returns today in exchange for the possibility of higher AI-driven earnings tomorrow. The financing pressure extends beyond individual balance sheets. Palihapitiya projects that the funding requirements of these companies through the end of 2027 may account for almost 30% of the anticipated issuance of new U.S. Treasury notes and bonds over the same period. While this is his estimate rather than an official government figure and should be treated with caution, the underlying idea is straightforward: many of the same large investors fund both corporations and the federal government. Increased corporate borrowing can crowd out demand for Treasuries, potentially forcing companies to offer higher returns to attract capital. That, in turn, raises the hurdle rate for AI investments themselves. The AI race is no longer solely about developing the best technology. Financing has become the central challenge. The companies once viewed as cash-rich engines of the economy are now testing how far their balance sheets—and the broader capital markets—can stretch to fund the infrastructure of the next technological era.

Anthropic’s IPO: The Tax Bill Nobody Seems to Be Talking About Everyone is focused on Anthropic’s valuation, revenue growth and IPO proceeds. I’m more interested in the stock compensation. Anthropic’s November 2 amended S-1 contains an important disclosure: the company expects to use IPO proceeds to meet tax withholding and remittance obligations associated with outstanding restricted stock units (RSUs). Why does this matter? Because employees don’t necessarily receive all the value of their stock awards in cash. When RSUs vest and settle, taxes become due. Someone has to provide the cash. There are several ways to handle this: Withhold shares to cover taxes. Sell shares through permitted sell-to-cover transactions. Use company cash to remit withholding obligations. Anthropic says it may use existing cash if IPO proceeds are insufficient. Now look at the bigger number. Reporting on the filing puts Anthropic’s unrecognized stock-based compensation expense at more than $32 billion as of September 30. That is NOT a $32 billion tax bill. It represents compensation expense associated with awards that have yet to be recognized. But it raises some important questions. How many shares will employees receive as awards vest? How much stock will be withheld or sold to cover taxes? How much additional dilution will shareholders face? How many employees and early investors will want to sell when lockups expire? Anthropic reportedly had $131.8 billion in cash, cash equivalents and short-term investments at September 30. So this isn’t necessarily a liquidity crisis. The company has substantial cash. But there is a difference between having cash, generating profits and paying employees with stock. Here’s what investors should be watching: Revenue growth gets the headlines. Stock-based compensation affects the actual economics of the business. An IPO creates liquidity, but it also creates an opportunity for employees and early investors to turn paper wealth into cash. The question isn’t simply what Anthropic is worth. It’s how much of that value belongs to existing shareholders after compensation, taxes, dilution and insider selling. I want to see the actual numbers before getting excited about the IPO. A $32 billion compensation obligation deserves more attention than another billion-dollar valuation headline.

Anthropic’s latest S-1 amendment is more interesting for what it reveals about the AI race than for any single headline number. Revenue reached $16.96 billion in Q3, up 47% sequentially. That still represents extraordinary year-over-year growth, but it is a sharp deceleration from the 144% sequential jump in Q2. The company swung back to a $567 million operating loss after a profitable prior quarter, largely because it leaned on more expensive short-term compute rather than cheaper multi-year deals. Price cuts and discounts already trimmed the quarter by roughly $600 million. Meanwhile, Anthropic has roughly $895 billion in computing and infrastructure commitments with partners including Amazon and Google, with tens of billions due in the next twelve months and nearly half of the total expected before the end of 2030. Some of that capacity depends on financing available to those partners. These figures sit against a simpler strategic question that most of the agent-launch coverage still skips: are the leading AI labs creating a new source of investment advantage, or are they competing to become the toll collector on information and tasks that everyone already needs? Toll-road businesses can be excellent. Payment networks and certain data platforms have proven that controlling a necessary intermediary layer, with high barriers and relatively stable demand once the capital is in place, produces durable returns. The model is less forgiving when the operator has already contracted for massive capacity before traffic volume and pricing power are locked in. If base models continue to converge and token prices keep falling, higher usage (the Jevons-paradox outcome) does not automatically produce enough revenue to cover the fixed commitments. The $600 million revenue impact from price reductions in a single quarter is an early signal of that tension. What matters more than the number of agents shipped is where incremental value actually accrues. Does the company possess proprietary data or evaluation advantages that improve with scale in ways competitors cannot easily copy? Are there meaningful switching costs once customers embed the models into workflows? Or is the primary position simply being the default API layer that customers pay until a cheaper alternative appears? The filing shows strong expansion in higher-spend customers and large remaining performance obligations, but it also shows the cost of locking in capacity ahead of that usage and the sensitivity to pricing. Those variables determine whether the economics look closer to a new moat or to commodity infrastructure with high operating leverage. Sequential growth rates will keep moderating simply because the base is larger. The more useful research is tracking whether the combination of usage growth, pricing power, and contracted capacity produces sustainable returns, or whether the industry is in a race to build roads whose traffic and toll rates remain uncertain. The Anthropic numbers make that distinction concrete rather than theoretical.

I didn’t discover AI when ChatGPT launched. I’ve been working in AI for more than eight years. My first AI company was AllSight. We built an AI-driven customer intelligence platform designed to help companies understand their customers by connecting and analyzing enormous amounts of data. In 2019, AllSight was acquired by Informatica. It was an important milestone. But there was another lesson. The technology didn’t work the way we needed it to. The product wasn’t scalable enough. The economics and complexity of delivering the vision were difficult. Having a great idea and building a commercially viable, scalable AI business are two very different things. I learned that firsthand. That experience shaped how I look at the AI industry today. I’ve seen the promises before. I’ve seen the demos. I’ve seen companies confuse an impressive technology demonstration with a product customers will actually use. And I’ve learned that the hard part isn’t getting AI to produce an answer. The hard part is making it reliable, useful, repeatable, scalable and economically viable. Eight-plus years in this business have given me a perspective that doesn’t come from reading the latest AI announcement or experimenting with a chatbot. I’ve lived through building it, selling it, and learning what doesn’t work. Today, everyone wants to be an AI expert. I’ve been paying tuition in this industry for years. And some of the most valuable lessons came from getting it wrong.

Anthropic’s IPO filing: $17 billion in quarterly revenue. $895 billion in infrastructure commitments. And growth is slowing. The amended S-1 filed November 2 gives investors a much better look at the economics of the AI business. Start with revenue. Q1 2026: $4.73 billion. Q2 2026: $11.56 billion. Q3 2026: $16.96 billion. Revenue grew 47% sequentially in Q3. Impressive, until you compare it with Q2, when revenue grew 144%. Year-over-year growth remains approximately 1,127%. The business is growing. But the rate of growth is slowing dramatically. Now look at profitability. Q2 operating income: $974 million. Q3 operating loss: $567 million. Q3 adjusted operating income was $2.46 billion, but adjusted operating margin declined from 19% to 15%. Compute and infrastructure expenses reached $9.86 billion in Q3, up from $5.56 billion in Q2. That’s a 77% quarterly increase. Price reductions and contractual discounts also reduced Q3 revenue by approximately $600 million. Customer adoption is real. Anthropic reported: 11,900+ customers generating at least $100,000 annually. 1,900+ generating at least $1 million. Approximately 250 generating at least $10 million. Approximately 900% net revenue retention, using its own methodology. This isn’t simply a consumer chatbot experiment. There is genuine enterprise demand. But here’s the other side of the filing. The infrastructure commitments. Anthropic disclosed approximately $895.4 billion in purchase commitments and leases, substantially related to compute and infrastructure. Of that: $447.5 billion is expected before the end of 2030. $69.8 billion is due within the next 12 months. Much of the remainder extends into 2031–2045. These are long-term commitments, not $895 billion of immediate debt. Anthropic has $131.9 billion in cash and short-term investments, plus $14.8 billion of available credit. But the company needs to keep turning infrastructure investment into revenue. Here are the questions investors should be asking. What happens when revenue growth slows? What happens when enterprise customers negotiate lower prices? What happens when compute costs rise faster than revenue? What happens when competitors offer comparable intelligence at lower prices? And how much additional infrastructure will be required to stay competitive? Anthropic has real customers, real revenue and impressive adoption. It also has a business model requiring extraordinary continuing investment. We have spent two years measuring AI by model capabilities, benchmarks, tokens and infrastructure announcements. The next phase should be measured by something much less exciting: How much cash does each dollar of AI infrastructure investment actually produce? Revenue growth is impressive. Cash returns are what matter.

The airport scale. You step on it. It records your weight. A number appears on a screen. And then what? Does the airline change your ticket price? Does it adjust the aircraft’s fuel calculations? Does it improve safety? Does anyone actually do anything with the information? Or did you just become another data point in somebody’s database? That’s the AI economy in miniature. We are collecting billions of data points, feeding them into increasingly expensive machines and celebrating how much information we can process. But the real question remains: When does the data change a decision? When does that decision save money, make money or prevent a loss? Until then, we’re just standing on a very expensive scale.

That’s actually a great analogy for the entire AI industry. You just got weighed at the airport. The airline collected the data. Your weight is a number. It can be stored, analyzed, compared and fed into an AI model. But here’s the question: When does that data actually matter? Does it change the price of your ticket? Does it improve aircraft safety? Does it reduce fuel consumption? Does it help balance the aircraft? Does it make boarding more efficient? If the answer is no, congratulations. The airline has another data point. That’s the problem with much of today’s AI economy. Companies count prompts, users, tokens, agents and dashboards. But the customer wants something much simpler: Did it save money? Did it make money? Did it reduce risk? Did it improve a decision? Collecting data is not creating value. Analyzing data is not creating value. Even using AI is not creating value. The value appears when something changes in the real world.

SNOWFLAKE: WHEN DOES THE AI GROWTH STORY BECOME A PROBLEM? Snowflake has a real business. It has enterprise customers, growing consumption and a rapidly expanding data platform. Its AI products are attracting attention. But there is another side to the story. Snowflake has grown revenue into the billions without establishing sustained GAAP profitability. In fiscal 2026: Revenue: $4.68 billion GAAP net loss: approximately $1.33 billion GAAP operating loss: approximately $1.44 billion Operating cash flow: approximately $1.22 billion Free cash flow: approximately $1.12 billion The difference between GAAP earnings and adjusted profitability matters. Stock-based compensation is a major part of that difference. Now consider the AI opportunity. Snowflake says its AI products are gaining adoption. But it does not separately disclose how many dollars of revenue AI generates or how much incremental profit those products produce. That leaves some important questions. 1. Does AI create new spending? Or are customers simply shifting existing workloads into new products? 2. What happens when computing gets cheaper? If AI models become more efficient, customers may need fewer computing credits to accomplish the same work. 3. What happens when competition increases? Databricks, Microsoft, Google and Amazon all want the same enterprise workloads. 4. Can Snowflake turn growth into GAAP earnings? Adjusted profitability is not the same as profitability after all compensation costs. 5. What happens if consumption slows? Snowflake’s business depends on customers continuing to use more of its platform. Its 126% net revenue retention rate is encouraging. But that number deserves watching if customers begin optimizing spending. Here is where I see the risk. Snowflake doesn’t have to run out of customers to disappoint investors. It can keep growing revenue, keep announcing AI adoption and keep winning enterprise accounts. But if the cost of delivering that growth remains too high, shareholders may never receive the economics they expect. AI usage is not profit. Revenue growth is not shareholder value. Adjusted earnings are not GAAP earnings. I want to see the dollars AI contributes, the incremental margins it produces and the actual earnings left for shareholders. Until then, Snowflake is a real business with an important unanswered question: Can the growth ever become durable GAAP profitability?