Sridhar Ramaswamy, CEO @Snowflake: Deepseek is Not a Threat to OpenAI & OpenAI Beats Anthropic|E1258
Summary
The most exposed AI startups build directly on foundation-model providers whose application boundary keeps moving. Sridhar Ramaswamy calls building on OpenAI “terrifying”: OpenAI, Anthropic, Microsoft, or Google can enter any promising coding, legal, or workflow category. Defensibility instead requires established customer relationships, clear delivered value, and embracing AI fast enough that a disruptor cannot unseat the incumbent.
DeepSeek may puncture claims about model scarcity without necessarily displacing ChatGPT’s consumer distribution. Harry Stebbings pushes back that DeepSeek reached No. 1 in the charts and is free. Sridhar answers, “It’s a product. It’s not a model”: ChatGPT bundles image creation, uploads, and code execution. Harry also suggests OpenAI could host DeepSeek to power part of ChatGPT.
OpenAI’s moat is approaching consumer-platform scale, not permanent model supremacy. By some accounts, ChatGPT has roughly 500 million loyal users without really paying for advertising in recent years; Sridhar compares that reach with Meta and Google. OpenAI may not always build the best or cheapest foundation model, but on the consumer side he bets specialized-model value will accrue to the incumbent entry point.
Established software companies can defend themselves if they combine embedded relationships with rapid self-disruption. Snowflake’s wager is that an AI-native entrant starting from zero will not be better than Snowflake, while Salesforce’s Agentforce illustrates the same playbook. “All-new value creation,” however, looks “very murky” where those advantages are absent.
Enterprise AI is producing real utility now, although adoption should be gentler than a frictionless hockey stick. Sridhar cites compressing notes from 30 Davos meetings—25 pages—into one-line summaries, querying structured data conversationally, and describing how parts of building-insurance underwriting could be automated by combining structured and unstructured information. The CEO message he heard was: “Help us create utility; tell us what is possible.”
The AI-capex arms race will end with a bubble bursting, but the residual value depends on what gets built. Harry points to Meta’s $65 billion data-center investment and the $500 billion Stargate announcement. Sridhar distinguishes a productive 1990s-style bubble that leaves power, buildings, and fiber from a Webvan-style burn—or rapidly depreciating hardware whose value disappears “in a puff.”
Snowflake accepts public-market constraints because accountability can sharpen innovation and expose narrative games. Private Databricks can spend more freely and has doubled the number of Snowflake’s salespeople, but Sridhar argues that constraints force clarity: “Having rich uncles is not always a good thing.” Public liquidity, free-cash-flow reality, and having to “show your work” outweigh the temptation to go private.
Deep dive
1. AI rewards malleability, while scale punishes yesterday’s strengths
Sridhar’s advice to a 21-year-old is to find work they care about that society values, then combine “drive and malleability.” AI is real, so the winning posture is to embrace change, invest in mastery, and remain nimble about where the opportunity moves.
Every knowledge profession will be affected because AI is an “incredible translation layer” between structured and unstructured knowledge. Software engineering might become an apex profession or a narrower one—journalism and music narrowed after internet distribution—but Sridhar considers disappearance unlikely while software keeps entering new domains.
His scaling rule is severe: whenever a team doubles, the traits that made its leader excellent often become “massive inhibitors.” A job suitable for 20 engineers can suddenly require 100; the leader is not necessarily incapable, but the business cannot wait while they reinvent themselves.
Hard conversations worsen when postponed and withholding feedback does the other person a disservice. Sridhar opens with humility—“This is going to be a difficult and unpleasant conversation”—then states expectations directly. His broader leadership tension is “driving while still being a good human being”; his separate parenting formula is “90% presence, 10% luck.”
2. Model providers can redraw the application boundary overnight
Application investing is murky because infrastructure and applications have become “super blurry.” Once a coding assistant or legal-document workflow takes off, there is no guarantee OpenAI, Anthropic, Microsoft, or Google will not build the same product; their motivation makes adjacency risk more than hypothetical.
The safer incumbent pattern is an existing customer relationship, clear delivered value, and the willingness to adopt AI before a disruptor can unseat it. Snowflake helps collect and analyze enterprise data, while Salesforce is pursuing self-disruption with Agentforce. Both can move from an established base while a newcomer starts from zero.
Harry’s sharpest pushback is DeepSeek: it reached No. 1 in the charts quickly and costs nothing, so why assume loyalty? Sridhar distinguishes model from product: ChatGPT adds images, file uploads, code execution, and other integrated capabilities. Harry suggests OpenAI could “shamelessly” host DeepSeek itself if that improved the experience.
OpenAI’s estimated half-billion users are the real strategic asset. Sridhar credits both the product and Sam Altman’s “incredible publicity machine,” while acknowledging DeepSeek exposed some mystique and misdirection around supposedly hard model problems. He nevertheless agrees with Altman’s warning: “It is terrifying to be a startup building on top of OpenAI.”
3. Snowflake’s moat must survive both giants and Databricks
Asked whether Nvidia could move up-stack, Sridhar broadens the threat to every cloud giant—AWS Redshift, Microsoft Fabric, Google BigQuery, and Oracle. Snowflake must fear being “a little mouse in the land of giants”; if they sneeze, it could be blown away.
His rebuttal is that product-market fit is a “living, breathing thing,” not something capital can summon. OpenAI and Anthropic, not Microsoft, Amazon, or Google, produced the leading models for much of the past three years; “money doesn’t buy you amazing foundation models, doesn’t buy you Snowflake.” Databricks remains a credible player, so copying is no substitute for continued execution.
Sridhar concedes that Snowflake was in catch-up mode in machine learning, but separates that field from post-ChatGPT AI. He claims Snowflake is now ahead on some fronts: data transformation, unstructured-to-structured extraction, and the forthcoming agentic framework Snowflake Intelligence, with Amazon, Elevance, Bayer, and other marquee names using Snowflake AI in production.
Private status lets Databricks “buy” business, ignore free-cash-flow pressure, and maintain twice Snowflake’s sales headcount. Public constraints instead forced Snowflake’s AI team to choose a focused agenda and catch up with modest investment; market scrutiny can overreact quarter to quarter, but liquidity and objective grading make self-deception harder.
4. Enterprise AI already works, but adoption will be gentler
Sridhar rejects both the ROI drought and a perfectly smooth exponential curve: growth should be gentler, yet AI already creates enduring value. After 30 Davos meetings, he put 25 pages of notes into Claude and got concise one-line summaries, while internal chatbots answered questions that would otherwise require clicking through dashboards.
CEOs were not broadly skeptical; they wanted vendors to “help us create utility.” The richer pitch combines document corpora and structured records inside agentic systems—for example, bringing together every relevant input so parts of building-insurance underwriting can be automated. When Sridhar described that possibility, CEOs responded: “Oh my God, that is amazing.”
Incumbents are moving faster because they remember repeated platform shocks. Companies like DEC disappeared, and SGI’s buildings were taken over; Facebook abandoned mobile web for native apps, while Google raised mobile monetization from 10% of desktop to 100%. Meta’s willingness to move on after augmented reality fell flat and redirect toward AI reflects learned institutional paranoia, not newfound startup DNA.
5. The capex bubble will leave either infrastructure or wreckage
Harry frames an unprecedented arms race through Meta’s $65 billion data-center investment and the $500 billion Stargate announcement. Sridhar’s endpoint is categorical—“a bubble bursting, as all bubbles do”—but the aftermath is unknowable: investment in power and buildings could create useful infrastructure, like the fiber laid during the 1990s bubble, while rapidly depreciating hardware could make capital disappear “in a puff.”
That uncertainty does not eliminate application opportunities because OpenAI cannot possibly handle every workflow. Sridhar offers Harvey as a niche he does not expect OpenAI to disrupt; the investor’s task is to identify similarly valuable niches that remain ripe for disruption.
Snowflake’s growth plan is primarily product-led. It is expanding beyond the analytics “gold layer” into ingestion, data engineering, machine learning, AI access, and Snowflake Intelligence—widening the addressable aperture rather than purchasing an unrelated extra ten points of growth.
Sridhar rejects turning Snowflake into “a PE shop,” though focused acquisitions remain valid; he says Snowflake spent roughly $150 million on Neeva and that it has more than paid for itself. The larger opportunity is customer-built data applications: Snowflake becomes part of partners’ revenue rather than merely an expense—“We make money when you make money.”
6. Distribution will decide consumer AI while enterprise stays fragmented
Google’s search dominance was not “immaculate conception.” It deliberately made deals for default placement through Yahoo, AOL, Firefox, and PC manufacturers while Microsoft slept; that gravitational center then supplied the traffic and leverage to defeat specialized search products.
Live.com once had better image search, but Google’s Universal Search placed images directly on the main results page. The same central entry point absorbed shopping, video, and maps—an instructive precedent for how an integrated AI product could capture value created by specialized underlying models.
Sridhar would bet that, on the consumer side, specialization will accrue to ChatGPT, which he thinks is becoming the entry point despite Harry’s skepticism. Enterprise lacks an equivalent universal front door after roughly 50 years of software fragmentation, so he expects multiple specialized models and applications—while retaining some concern that a single gateway might still emerge.
Being first mattered “100%”: Google reportedly paid AOL more than it made from the deal—over $100 million per year—to secure a key distribution entry point. Sridhar’s enduring lesson from Larry Page and Sergey Brin is operational relentlessness: every argument was exhausting, but pushing first principles through Google Books, YouTube, and the Dutch-auction IPO repeatedly produced unconventional business advantages.