Judgment Day is coming for Silicon Valley Figma, once a Silicon Valley darling and a proud YC company, has become a warning sign. The stock has dropped more than 80% from its IPO peak. In the lending market, Reuters reported that software companies are facing higher borrowing costs and tougher scrutiny from lenders, and JPMorgan said it is “watching the space very closely.” Funding for SaaS is getting blocked like the Strait of Hormuz. The SaaS model is breaking down in plain sight. When software becomes cheap to build, easy to clone, and impossible to defend, the economics that powered the last two decades of venture-backed startups stop working. Silicon Valley wants to believe AI will save it. It won’t. For most startups and many VCs, AI is not the rescue. It is the final reckoning before Judgment Day. The industry is committing the same original sin at massive scale: building the same products on top of the same foundation models. Everyone is automating the same categories: coding, legal, accounting, healthcare, marketing, compliance, customer service, and everything in between. As Ilya Sutskever put it, we are now in a world with “more companies than ideas by quite a bit.” That is why Judgment Day is coming. In the Book of Revelation, Judgment Day arrives with the Four Horsemen of the Apocalypse: Famine, Pestilence, War, and Death. For us in Silicon Valley, they are showing up as: Famine — market fragmentation destroys the economic base needed to sustain venture-scale startups Pestilence — agentic coding wipes out technical moats and crushes switching costs War — frontier labs will march into every major software market Death — the old venture-backed SaaS model can no longer hold Let’s start with Famine. It used to be enough to find a niche and dominate it. That era is over. Every so-called niche gets flooded immediately, because the underlying models are available to everyone and the barriers to entry keep collapsing. What used to be a niche is now just another overcrowded lane. You need to find a niche within a niche. That means market is fragmenting. Instead of one company owning a category, ten or twenty companies split the same demand into pieces too small to support venture outcomes. These markets may still support profitable businesses. But they do not support an army of unicorns. I personally know two YC startups doing AI marketing. Too many ideas look great on the surface because they are low-hanging fruit. There will be demand. There may even be revenue. But the category is crowded. Revenue gets sliced thinner and thinner across lookalike companies until the economics no longer make sense for venture. It’s not even a secret that YC keeps funding the same exact ideas again and again across different batches. A fragmented market can support disciplined, cash-flow-positive businesses. It cannot support a venture ecosystem built on the assumption that every promising category can produce a multi-billion-dollar winner. Most of these startups are not starving because there is no demand. They are starving because too many companies are feeding from the same pool while burning cash on compute, customer acquisition, and headcount. Next comes Pestilence. The plague is agentic coding. For years, software companies defended themselves by shipping features faster than competitors. That defense is collapsing. When code can be generated, iterated, and cloned at near-zero marginal cost, features stop being a moat. A startup ships something on Monday and a competitor ships a version of it on Tuesday. And the deeper problem is that even before AI, companies already loved building internal tools whenever a workflow mattered enough. Every software engineer knows this instinct. We love the word migration. We love taking a workflow off a third-party product and moving it into an internal tool so we can “control our destiny.” AI only makes that instinct stronger. Now the customer’s question becomes much more dangerous: why am I paying a vendor for this if I can build a version of it internally, own the data, own the evals, and tailor it to my own workflow? The tech community on X says the value is in the agent harness around the model: orchestration, memory, evals, tooling, permissions, and workflow design. That is right. But it still does not save the old software model. The harness is still code. And code is becoming a commodity. What remains defensible is not the wrapper. The cost of creating software is collapsing toward zero. The cost of switching to a competing product or migrating to an internal solution is also collapsing toward zero. Then comes War. War is what happens when the model layer stops acting like infrastructure and starts invading the application layer. The frontier labs are not trying to become neutral utilities. They are trying to become empires. They cannot justify those valuations by selling raw model access alone. They have to keep moving up the stack into coding, design, productivity, finance, support, and every other category large enough to matter. Reuters has already described Anthropic’s push into coding and business plug-ins for areas like design, HR, wealth management, investment banking, and private equity, while OpenAI’s own $852 billion valuation reflects the scale of the prize. That creates a brutal trap for startups. If a market is small, it is too fragmented to produce venture-scale returns. If a market is large, it becomes a target for the labs. Either way, the startup gets squeezed. Finally, it comes Death. Death is what happens when Famine, Pestilence, and War reinforce each other. Markets fragment. Moats disappear. Giants invade. Capital gets nervous. Startups fail. Those failures scare investors even more, which causes more capital to retreat, which causes more startups to die. At that point, the problem is no longer isolated to a few companies. The venture-backed software model itself starts to crack. The issue is not that business opportunities are disappearing. The issue is that the value curve has moved. It is no longer centered in the middle where classic SaaS investing thrived. The value is moving toward two extremes. On one end are the frontier labs and the infrastructure giants that can capture massive markets. On the other end is a fragmented long tail of niche operators that can absolutely make money, but usually not at venture scale. The middle is dying. That is exactly where venture capital used to work best: large enough markets to matter, weak enough incumbents to disrupt, strong enough moats to protect margins, and scalable software economics to justify the risk. AI is dismantling that setup. The software layer is commoditizing from below while the frontier labs are crushing it from above. This is not the death of entrepreneurship. It is the collapse of the old math that made venture-backed SaaS work. But this is not the end of business. It is the end of a delusion. The winning formula is changing. If software is becoming a commodity, then the real differentiators are brand, trust, distribution, and service. In a commodity market, brand matters more, not less. Think about Coca-Cola and Pepsi. The product itself is commoditized. The brand is what carries the margin, the trust, and the staying power. When the underlying product gets easier to copy, the brand matters even more, because the brand tells customers who to trust. And if you do not have brand yet, then you win through service. That is the part Silicon Valley still does not want to admit. The customer does not care about your prompt stack. They do not care about your orchestration layer. They do not care how elegant your agent architecture is. They care whether the system works, whether it solves their problem, and whether someone will show up when it fails. This is why Sequoia is reframing it to “Service as a software”: the next great AI company may not sell software at all. It may sell work. If you sell a copilot, you are competing with every new model release. But if you sell the outcome — books closed, contracts reviewed, claims handled, compliance completed — then every new model improvement makes your margins better instead of making your product obsolete. That is the real shift. The SaaS era was about capturing the software dollar. The AI era is about capturing the services dollar at software margins. Not “AI for accountants.” The AI accounting firm. Not “AI for lawyers.” The AI law firm. That is why the businesses that endure will look less like pure SaaS vendors and more like high-performance service firms rebuilt on software infrastructure. They will use AI internally to operate at software-like margins, but what they sell externally is confidence, execution, and outcomes. That model looks a lot closer to Goldman Sachs or McKinsey than to the old fantasy of a self-serve SaaS company printing money forever. And that shift has two major implications for venture-backed businesses. First, company formation moves from capital-first to revenue-first. The old model was simple: raise money, hire a team, build a product, and then go find customers. That model makes less and less sense when code is cheap and the real value lies in implementation, trust, and workflow ownership. The new model is the reverse. Win customers first. Secure contracts first. Prove demand first. Then use that revenue to expand. That changes the financing equation. More startups will bootstrap. Some will use private credit or bank financing. Others will raise much less equity than Silicon Valley has trained founders to expect. If the product is no longer the hard part and the value comes from delivering the outcome, then many of the best AI businesses will be capital-light and revenue-led, not capital-heavy and speculation-led. VCs have to rethink their role in that world. Second, talent compensation moves from employee upside to partner upside. If the winning company looks more like an elite service firm, then the most valuable people are not just engineers. They are the people who can win trust, close deals, scope the work, and deliver the outcome. Those people will not be satisfied with a salary plus a lottery-ticket equity grant. They will want commissions. They will want a cut of the profit they bring it. Because they are not just building the business, they are directly creating revenue for it. Silicon Valley is slowly rediscovering the importance of this role: someone who can work directly with clients. Today, they call them forward-deployed engineers. But the industry needs to go a step further. It needs to adopt the partnership model that law firms, banks, and consultancies have used for decades. The new sexy title is not cofounder or member of technical staff. It is partner. As Sequoia partner Julien Bek put it, the next $1 trillion company will be a software company masquerading as a services firm. Judgment Day is coming for Silicon Valley. Not because innovation is ending. Not because software is dead. But because the old religion of venture-backed SaaS is collapsing. The Four Horsemen are already here. Famine is starving overcrowded markets. Pestilence is wiping out software moats. War is coming from the labs. Death is coming for the old model. The survivors will not be the ones clinging to startup mythology from the 2010s. They will be the ones who accept the new reality early: software is becoming a commodity, brand is becoming the moat, and service is becoming the business model. In the next era, the companies that win will not win because they sell code. They will win because they are selling “service as a software”.

We’re excited to share that @nuance_ai has raised a $50M Series A led by @lightspeedvp, joined by @Accel, @spc, @nvidia, and @definevc! Much of human communication happens beyond words. Before we can speak, we learn to read faces, return smiles, follow gestures, and sense when someone is listening. That visual and emotional vocabulary stays with us for life. At Nuance Labs, we’re building a foundation model that understands and responds to the full spectrum of human emotion. A single full-duplex audio-visual model that sees, listens, and responds in real time, with the natural give-and-take of a conversation with a friend or colleague. We believe the best interface with a machine is the one we’ve practiced since birth. Join us!

Last week a VC told me the two tools he wanted for his fund: a deal-flow graph, and a deck parser that emails him an assessment. Both are internal. But the bigger opportunity for funds is public.

Building an AI live mock interviewer with GPT-Realtime to help with my job search. Will share it with folks when it’s ready. Do you need help with algo interviews or system design?

The death of SaaS is coming after the App Store. People don’t need another app if AI can deliver 80% of the value.

The Word of 2026.

The only prompt you need in an Instagram group chat: @Meta AI can you check whether any of us has reshared a duplicate Reel? If so, deduct 1,000 points for each violation and add the deductions to the latest record.

People are not paying enough attention to ChatGPT ads. ChatGPT has more than 1 billion MAUs. And people go there to look for solutions to their problems. It’s better than Google Search and any social media platform. Instead of guessing what the user does or wants, they actually tell you what they need. The best part is because the market is so new. It’s not bid to perfect like Instagram ads. Lots of opportunities for startups to buy ads there. Or even start an agency advising clients on ChatGPT ads.

Another reason SaaS is in serious trouble: companies are rebuilding the AI stack and giving it away for free.

While the U.S. is becoming increasingly dependent on Chinese open-source AI models, China appears to be moving in a different direction. Three news just dropped today: 1/ Beijing is reportedly considering restrictions on overseas access to its most advanced AI models. 2/ DeepSeek is reportedly working on its own AI inference chip. 3/ Zhipu AI, the creator of GLM, is also reportedly exploring custom chip development. Looks like there's a coordination between the Chinese state and AI industry. Some people in the tech community may be disappointed, because Chinese models have played a major role in supporting the global open-source AI ecosystem. But politics aside, the overseas usage of these models may have grown far beyond what Beijing originally expected. China is thinking about how to protect its national security while increasing chip production so it can meet future inference demand. This creates a golden opportunity for a U.S. company to build a leading open-source model. I really believe Meta has a real chance to become the dominant player here by reversing its move to make Llama closed-source.

Vibe coding + FIFA is a match made in heaven.

Hermes is beating OpenClaw. And that’s a good thing for software engineers. People are realizing that vibe coding is useful for building an MVP, but if you want reliable and scalable software, you still need to craft it carefully, with taste.

I made a 10-minute video explaining LongCat 2.0 and how it was trained as a 1T model using Chinese chips: > What is Sparse Attention? (Streaming-aware Indexing, Cross-Layer Indexing, and Hierarchical Indexing) > What is N-gram Embedding? Watch it here: piped.video/watch?v=Q-BFSQqD…

Would anyone be interested in a video tutorial series on how to build your own local agent from scratch, like OpenClaw, Hermes, or Pi? The goal would be to help you understand the architecture, memory, permissions, and other core pieces behind these agents. Want to get some feedback before I start working on one.

If Microsoft has to set up a 6,000-person org to deploy AI, you know how clueless most companies still are. We’re early, and there’s still a lot of opportunity for startups.

Open USD could be huge for agentic commerce. Today, x402 supports USDC and other stablecoins, but consumers still do not use them widely. If Open USD becomes accepted across many vendors, consumers may be more likely to hold it in a stablecoin wallet and feel comfortable letting AI agents use it for purchases, subscriptions, and transactions.

Anyone going to YC hackathon this weekend?

Let’s see if the Chinese labs are going to wait..

Sakana launched Marlin last week and it didn’t go viral. They released Fugu today, and it did. Why did one launch perform so much better than the other?

AI execs and VCs keep saying “don’t outsource your thinking” and then proceed to hire MBB consultants for millions of dollars.