Transcription
On February 25th, Perplexity shipped the best Agentic product of the month, and it might not even matter. Perplexity Computer launched to genuine excitement. I've tried it. It's a cloudnative multimodel orchestration system that routes work across 19 different Frontier models. It spawns sub agents. It persists for months and it delivers finished artifacts while you sleep. It sounds amazing and it's really, really good. It runs Claude Opus 4.6 six as its reasoning core, Gemini for deep research, Grock for speed, chat GPT 5.2 for long context recall. It is available today for 200 bucks a month and it's almost certainly worth it if you do heavy research and ops workflows and I'll get into those.
But it is also structurally a cautionary tale about where most of the AI industry is building right now and why good execution on the wrong layer of the stack will not save you. I want to be really clear here. Perplexity launched a great product. This is absolutely not a Perplexity hit piece. Perplexity is one of the bestrun AI companies in the world. They read the market correctly. They made a genuinely bold call to kill their own ad business in February because they understood that trust is the new distribution and they are now targeting $650ome million in 2026 revenue. Their search API already has four of the Magnificent 7 running it in production.
And yet and yet their core reasoning engine runs on a direct competitor's model. Their deep research runs on yet another competitor's model. Their speed layer runs on a third. Every model provider they depend upon is simultaneously building the exact product computer competes with the weak perplexity launched computer anthropic ship. The enterprise expansion of claude co-work with deep connectors private plug-in marketplaces and the ability to pass context seamlessly across tools. Co-work doesn't need 19 models. It has one and it owns it.
So that asymmetry between the quality of Perplexity's execution and the fragility of its structural position, that is the thing I can't stop thinking about because it's not just Perplexity's problem. It's the position of almost every AI company that isn't anthropic, Google, OpenAI, or Meta. So the question I want to dig into with this video is simple. How do you diagnose whether you're building durable position or just renting it in the market? And if you're renting it in the market, what do you do about that? And yes, along the way, we're going to talk about Perplexity Computer, cuz it really is a great product.
But before we start, I want to talk about how February 2026 redrew the map for AI. I've talked about this before, but it's worth emphasizing again. The first few weeks of 2026 have unlocked patterns that will take years to play out. Let's start with a basic sequence. In late January, OpenClaw, an open source AI agent built by Austrian developer Peter Steinberger, hits a 100,000 GitHub stars. It's well over 200,000 now, by the way. And it becomes the fastest growing repository in GitHub history. It runs locally. It connects through WhatsApp and Telegram. And it takes autonomous action. It manages email. It modifies files. It browses the web. The demand signal is massive. The trust problem is equally massive. Cisco's security team finds a third-party skill of performing data exfiltration and prompt injection without user awareness. Someone's agent deletes all their emails. Anothers creates a dating profile and starts screening matches without being asked. Imagine that. But despite all of the security flaws, people keep showing up on Moss and OpenClaw eventually spawns an entire project that Peter is now leading at OpenAI.
Meanwhile, Enthropic launches Claude Co-work on January 13th in research preview. It's since expanded. This is described as clawed code for the rest of your work, which sounds big, but is probably underelling it. February 5, just a couple of weeks later, Tropic ships Claude Opus 4.6 with a million token context window. Within days, iShares expanded tech software ETF records its worst 2-day stretch since 2008. I'm not going to belabor this. You've already seen it. There's seven or eight selloffs in SAS stocks. There's something like a quarter of a trillion dollars wiped out of the markets. Fundamentally, people are looking at co-work and seeing SAS blood on the walls, but Anthropic just keeps shipping. On February 11th, co-work ships on Windows. 70% of desktop computing now has access. February 14th, Valentine's Day, Peter Steinberger officially joins OpenAI. So, that OpenClaw project starts to get sponsored by OpenAI. And Peter has the resources now to really build out a secure OpenAI sponsored version of OpenClaw. February 18th, Perplexity confirms that it has abandoned advertising and tells the Financial Times that ads risk making users lose trust. February 24th, Anthropics Enterprise Agents launch with deep connectors, private plug-in marketplaces, and pre-built templates. Are you out of breath yet? There's so much here. February 25th, Samsung unveils the Galaxy S26 Aentic AI phone. By the way, Perplexity is running on that. Google previews Gemini agents with a framework called App Functions. the same day that lets apps expose data directly to AI agents at the OS level and oh by the way also perplexity ships computer that's a lot so in one stretch for about 6 weeks between the first half of January and the end of February the industry stratified into layers with fundamentally different structural economics at the bottom the model providers own the weights in the middle orchestration and application layers combine models into products at the top distribution owners control the surface where users encounter agents. Hovering over everything are cloud providers spending $690 billion a year on infrastructure they must fill with tokens.
Now that structure was something that you could see on the wall in 2025. All that January and February did was harden the demand signal and reveal who is playing at multiple levels simultaneously. The problem for perplexity is that perplexity sits in the middle layer which is the most exposed layer in the system. When a technology stack consolidates, the layer that gets squeezed is the one between the platform owner and the customer. It happened to travel agents, it happened to media companies, and it happens to enterprise middleware. The common thread is if you don't own the layer below or the relationship above, you're just borrowing time. In AI, you're in that trap if you build on models you don't control and serve customers that model providers are now selling to directly. Every upstream provider has the ability and now the incentive to replicate what Perplexity does or to change pricing and access terms in ways that compress Perplexity's margins. For example, reports have surfaced that Anthropic began banning users who powered OpenCloud with claude credentials. Similar reports have surfaced about Google and OpenClaw. If that logic of coming for other players in the system extends to orchestration layers, the dependency risk for perplexity is not going to be theoretical. It's going to be very very practical.
But as bad as that is, the squeeze on perplexity isn't only coming from below. The same players squeezing perplexity from below are also coming for the context layer, coming from above. So the conventional defense for middleware companies is we have domain expertise. the modelmakers cannot replicate. Open AAI Frontier just blew a hole in that argument. Frontier launched as an enterprise platform that connects silo data warehouses, CRM systems, and internal applications into what OpenAI calls a semantic layer for the enterprise. The idea is to onboard agents with institutional knowledge to grant them identity and permissions and to build evaluation loops so agents improve with experience. That is the context layer. Now, it is not as fully realized as it could be. We're not talking about superhuman intelligence operating across a 10 token context layer, but that's where OpenAI wants to go with it. And that context layer, that's supposed to be your moat if you operate at the middle level. Harvey, Sierra, Decagon, and a bridge are already committed as frontier partners. So smart players who build on top of models are already joining forces with open AI because they don't want to get eaten.
Now, this doesn't mean that domain context is worthless. It means the form that survives may be narrower than people think. If your domain expertise is mostly connecting enterprise systems and teaching AI how your org works, Frontier does that now with forward deployed engineers. If it's something deeper, if it's proprietary data, if it's regulatory knowledge from years of compliance, if it's operational insight from running specific physical processes, there's still value there. But most companies claiming domain modes have not done the rigorous thinking to figure out if they have true domain expertise that survives this kind of context consolidation or not. The thing to keep in mind here is that the hyperscalers are not neutral referees. They are coming for the tokens. They need trillions of tokens to justify their valuations and to justify the capital spend. It's just simple math. If OpenAI can't hit its target of 250 to $280 billion in a few years in revenue, then the entire capital structure of the system doesn't work. They have to do this. It's an existential bet. That is true at differing scales for every hyperscaler in the game. In that position, these are not companies that can afford to be neutral platform providers. Every layer of the stack that a hyperscaler controls can generate tokens in ways that benefit that hyperscaler's infrastructure bets. Every layer a hyperscaler does not control is a layer where somebody else captures value from compute that they're subsidizing. This is exactly why AWS secured exclusive thirdparty distribution for Frontier and is co-building the stateful runtime environment on Bedrock. That is vertical integration, guaranteeing the enterprise agent layer runs through its infrastructure. Microsoft is taking a 20% revenue share from OpenAI through 2032. In many ways, that's much simpler, but it also locks in their ceiling. Meanwhile, Google has invested $3 billion in Enthropic while building its own agent layer into Android. And Amazon backs both OpenAI and Anthropic, all while building its own custom tranium silicon. In many ways, the framing cloud providers win regardless is true, but remains an early 2025 story. The 2026 question is for those hyperscalers, which layer generates the most tokens that soak up demand and the hyperscalers are structurally compelled to own as many of those layers and the fattest layers they possibly can to justify their infrastructure buildout. That is the world Perplexity Computer launched in.
And now it's time to talk about why Perplexity Computer is special. First, despite the name, it's not hardware. It's a cloud-based Agentic system that orchestrates 19 AI models together to execute really complicated multi-step workflows end to end. It's right now available exclusively on Perplexity's $200 a month Perplexity Max tier. It represents the company's clearest bet yet that the value layer in AI is not the model, it's the orchestration. So the core idea with perplexity computer is that you describe an outcome and computer decomposes it into tasks and subtasks. It will spawn specialized sub aents that run in parallel. One agent might do web research while another drafts a document. A third generates visuals and a fourth writes code. Each task runs in an isolated compute environment with access to a real file system, a real browser, and integrations with tools like Gmail, Slack, GitHub, Notion, Salesforce, and more than 400 others. The model routing is what Perplexity is claiming is their differentiating architectural choice. Computer uses Opus 4.6 as its central reasoning engine and then delegates to specialized models per task. This routing is automatic, but users can override. You can pin specific models to specific subtasks if you have strong preferences or want to manage token budgets. Crucially, workflows can run asynchronously for hours or even months. So, you can kick off a job, you can close your laptop, and you can come back to finish deliverables. Computer retains persistent memory across sessions, so it accumulates context about your preferences and past work over time. This all sounds amazing and Perplexity deliberately positioned the launch as a secure responsible version of Open Claw.
But where are users actually using it and to what extent is the usage of computer tied into Perplexity's other core competency search? Well, the most credible early use cases are clustering around research heavy multissource workflows. For example, competitive intelligence and market research. Hey, that calls out search. Give it a prompt like analyze the top five competitors in X space. track their recent product launches and produce a briefing. Computer parallelizes seven different search types simultaneously, reads full source pages, deals with academic sources, deals with web sources, and it cross references findings and then constructs a structured detailed report. Another use case, financial analysis and investment memos. This also plays into something Perplexity has been investing in for a while. You can pull earnings data. You can compare margins across competitors. You can synthesize analyst sentiment. And you can output a formatted PDF with charts if you want or a website. This is where the multimodel routing really earns its keep. You can have research agents gathering data while the coding agent builds the visualizations. Another one that's a little more surprising is outbound and pipeline building. an early reviewer testing cold outreach automation where you have computer find real email addresses research each prospect's recent activity draft personalized messages referencing specific details and send them through a connected Gmail account the recurring version daily competitive monitoring weekly reports is where this starts to shift from a one-off assistant into a persistent agent that can help with outbound end to-end build tasks are something that Perplexity has also touted here like build me a personal finance dashboard or create a portfolio site for for me with case studies and then deploy it. Computer can handle the research, design, code, and deployment loop in one long session, although it currently stamps outputs with a watermark. So, you may not want to distribute it too widely. Last but not least, think about content repurposing. You could pull a segment from a long podcast, extract the clip, convert it to vertical format, and add captions. Essentially, it becomes a multi-tool pipeline that would normally require stitching together three or four services, and you could just do it with computer.
So when I look at that universe of use cases, I do see that a lot of them skew into what Perplexity traditionally does well, and that's not a surprise. My honest answer to who should pay 200 bucks a month for this is frankly power users in those specific verticals. If you already spend $100 to $200 a month across multiple AI tools and you're tired of context switching and any of those power use cases resonate for you, this may be a knock it out of the park win for you. Perplexity's own executives have been explicit that they're targeting people making GDP moving decisions, not just maximizing free tier monthly active users. So if you're a founder who's doing your own research, if you're a solo operator, if you're an executive, they are targeting you and they want to serve you better over time. This makes it a natural fit for consulting and advisory work where the deliverable is a synthesized research artifact. The parallel research engine is genuinely strong for academics as well. Early reviewers consistently note that the research capability, not the flashy code generation, is where computer overdelivers. This is my surprised face. It plays right into their search system. But if you are not in that category, if you are not a leader who wants to combine all of their AI usage into one place and wants to lean on research and wants to lean on consulting deliverables, maybe wants to go into that outbound thread or go into finance, you should probably skip this because if your workflow is primarily single model, if it's conversational or if you're doing deep technical coding, this is just not a fit for you. Computer's autonomy is a feature when it works and it can be a liability when you need precision on a nuance task. I would not be using computer if I was building a large or long development project.
So perplexity strategic position its multimodel orchestration is a moat that is not a great moat to be honest with you. It is easy to replicate that moat. So the risk for them longterm is the commoditization of the models themselves is going to make the orchestration layer much less valuable. There's also some practical issues with rollout. Perplexity's planned live demo was cancelled due to product bugs, a signal that reliability at this level of autonomy is not fully baked yet. So, buyer beware. Still, Perplexity Computer is the most ambitious attempt so far to package truly multimodel agentic AI into a tidy consumerf facing product and we should pay attention to it for that alone. We will see more work in this direction. The research engine and parallel execution, the model routing are all particularly strong. Persistent memory and 400 plus integrations gives you a chance to make real workflows that matter for you. Now, it's not cheap, right? At 200 bucks a month, they're willing to price for professionals. But if that's you, if any of these use cases have resonated, you should go ahead over to Perplexity Computer and give it a shot. I think you'll be surprised. And yes, I put together a guide on how to dig in and what to make of these use cases with some sample prompts over on the substack.
But having looked at Perplexity Computer in a little bit more detail, I think you're going to start to see where Perplexity is at risk as a company. Because if we step back into the middleware conversation we were having, Perplexity Computer does not solve any of Perplexity's problems. In fact, really, February of 2026 just revealed the trap that Perplexity is sitting in. Perplexity is a company struggling with structural durability, no matter how great their execution is. In that world, it's natural to ask yourself which structural positions in the software AI stack survive, which have durability, which have an edge. I want to suggest to you that there are four that do.
Position number one, know which context to platform, which to retain, and why. So, there are three kinds of enterprise context. Each of these has a different answer and yes, you do need to get into this level of detail to think through a structural edge in 2026. So the first type of context is structural context. How systems connect, where data lives, what permissions exist. This is commodity plumbing. Frontier's connectors are going to handle it. If your startup's moat is wiring up Salesforce to Jira, Frontier is going to eat you on a 12 to 24month timeline probably much faster. Another layer is operational context. How decisions actually get made. the informal knowledge in senior people's heads. That's more interesting. I think Frontier is aiming to approximate it, but the approximation does have limits at least at the moment. The rate of change here is the variable that determines your defensibility. If your operational context updates quarterly from regulatory filings, maybe the platform keeps up. If it updates hourly from live physical operations, the platform might not be able to. So for slow changing operations, this window closes faster. For companies that are generating novel data daily, you have a much wider edge to work against. Proprietary context is the other one to think about. Data and judgment that exists nowhere else where handing it to the platform would create competitive exposure like a trading desk's risk models, a drug company's experimental data, a manufacturer's sensor data from proprietary production lines. This is where the enterprises calculation flips because platforming it means giving the provider access to the thing that makes you you. So enterprise buyers need to ask does this give the provider what they need to compete with me? Can I get my context out if I switch? What are the regulatory constraints here? So if we look at position one, the middleware companies that survive are going to be companies where context value lives either in the proprietary layer. So enterprises don't want to give that away or the operational context is updating really really frequently and you just can't get that into a one-sizefits-all system. That is the context structural advantage that a middleware company can take and win. And that's just position one. We have four different positions you can win on. So if you think this is all about how middleware is doomed, it's more complicated than that. Middleware is a fragile place to be. But there are interesting opening niches that we're going to explore here. Now I want you to recognize Perplexity is not in position one. Computer did not get it there. Perplexity does not have this advantage.
Position two, middleware companies can become the infrastructure the agents call. This might be Perplexity's most durable business. Their search API already has four of the Magnificent 7 using it in production. And the search API powers systems like co-work infrastructure that agents need regardless of who wins the orchestration war is perhaps their most resilient play. This is part of why they partnered with Samsung on those Android phones. So Quarter, a financial data company, published a case study showing how Perplexity Finance uses their earnings transcript API as a core data layer. Quarter doesn't compete with Perplexity. It's just infrastructure that Perplexity cannot function without. That's the model. data feeds, verification services, domain specific APIs, compliance checking. These are the picks and shovels of the agent era. They're structurally safer than building another wrapper because the agent providers are your customers, not your competitors.
Position three, be the middleware company that owns the customer workflow deeply enough so that switching costs actually protect you. When Anthropic shipped private plug-in marketplaces where organizations deploy custom agents with institutional knowledge encoded into workflows, that's not a chatbot. That's a system that learns how this specific legal team reviews those specific NDAs. Ripping it out means rebuilding all of that with institutional encoding. So again, I've said this before, stop thinking about model selection. Start thinking about integration depth. The question isn't which model you use. It's how many institutional workflows break if someone rips your system out. Deep integration takes a whole lot longer than wrapping an API, but it produces something the wrapper approach can't. A product that gets more valuable the longer it runs. Perplexity is not in this business.
In position four, be the middleware company that owns the trust and verification layer. February of 2026 revealed three incompatible trust architectures that all deployed at the same time. Open claw local no enforced security perplexity computer in the cloud curated sandboxes and Google's Gemini agents OS level containment. As agents start to proliferate someone needs to audit what they do, verify their outputs and enforce policy. The model providers are very focused on capabilities right now. Maybe not so much on governance. So the gap between agents are doing real work and we can prove what agents did is wide and growing. This is analogous to what accounting firms did when financial complexity started to outpace regulation and we don't have a solve for this but it represents a massive building opportunity in the middleware space.
Despite these opportunities, I also want to be honest with you that there are places in middleware that look like deadends to me that look like corners where the hyperscalers will just eat. And I want to outline a couple of them to you today. In fact, three of them. First, which cloud runs your tokens is becoming a zero sum game because there are only so many tokens out there. And yes, I know tokens are being generated hyper exponentially, but there is still an upper limit that continues to grow. If you are in the game of running tokens on your cloud like AWS, every single token that AWS can secure is a token that does not go to Azure or Google Cloud. And critically, in a world where AI is scaling quickly, a given token running on AWS is incrementally influential of the next token. In other words, strategically speaking, capturing X share of the token load from an enterprise means you're more likely to capture more share because these things have a center of gravity effect. If State Farm ends up deploying Frontier agents, those tokens ipso facto go to AWS. That becomes a customer Microsoft doesn't get. That is a zero sumum game. I would not want to be a middleware company getting in the way of multicloud orchestration or trying to play in those waters. Second zero sum game. Second blind alley. Don't run down here. Which layer captures the margin? So token volume can grow 10x and the middleware layer can still get squeezed down towards zero if model providers absorb all of that functionality. Think about where the value is in the business. So when I outlined the four places where I see middleware still thriving, every single one of those focused on sustainable differentiated value. In other words, tokens that mean something. If you cannot show that your tokens add differentiated value on a sustainable basis beyond baseline vanilla tokens from model providers, your margin is going to get squeezed into nothing into zero inevitably. you will simply stop generating margin even if your service continues to exist. Third, do not get in the way of who owns the enterprise relationship. When OpenAI deploys forward deployed engineers alongside customers to set up frontier, it is building the relationship that determines which AI services enterprises buy for the next few years. Anthropic is doing the same with their forward deployed engineer. When a relationship is locked in through forward deployed engineers, a vertical startup has to fight hard to find space in the conversation. If you are a middleware company, what that implies is you need to find a corner that has clear and differentiated value where you can build a relationship with an internal stakeholder or champion who understands AI really fluently. You cannot be looking for the dumb people here. You cannot be looking for the people who are behind the ball. You cannot be looking for the people at risk of getting fired. You have to be looking for the people who are super AI fluent, who are champions, who are smart enough to understand that you as a middleware provider are building something structural that AI companies by themselves, hyperscalers by themselves are not going to go after and that therefore understands why you need to exist. I know that is a lot more strategy than we have traditionally had to put into the sales path and that demands more of our sales teams. But there is not another way to do this because word is starting to get out that these hyperscalers are eating everything in sight and suspicious CTOs are saying if we wait 6 months can't claude just do it if we wait 6 months can't open AI just do it. You got to have an answer to that as a middleware company and I've outlined the positions that I think you can play.
So stepping back what lesson does perplexity actually teach us here? Perplexity is simultaneously in the middleware trap and I've talked about that at length and is also building one of the most interesting ways out of that stack that we're not talking about. So everyone talked about computer. I had a whole little segment on computer in this video but people aren't talking enough about the fact that computer is not the way out of the middleware trap. The search API is agent infrastructure which Perplexity is also building that got much less press. That is their strategic out for the middleware trap. The decision to kill advertising to protect trust tells you leadership is starting to understand the structural game. The executives are very much signaling that they are in the accuracy business. They are in the high margin, highv value customer business and they are in the business of providing agentic infrastructure. That's a pretty smart play. The lesson for many other middleware companies is harder than it looks. Anthropic owns the model and is aggressively collapsing layers around knowledge work. OpenAI owns the model and is aggressively building an enterprise context platform. Google owns the phone, the model, the cloud, and the browser. Meta owns distribution and bought Manis for the execution layer. AWS and Azure are structurally compelled to own as much downstream token generation surface as possible. If you're not one of those companies I just named, and almost none of us are, your strategic imperative is not to compete with them directly. Please don't try. It's to find the position where their incentives align with your existence rather than your replacement. Infrastructure that agents call as a place workflow depth that generates tokens the hyperscalers need. Governance the ecosystem needs proprietary context the enterprise can't rationally hand to a platform provider. These categories are narrower than a lot of pitch decks tend to suggest, but they're real and the enterprises that need it tend to know they need it.
So I think this is what durable position looks like in 2026. The bar is higher than it was. The window to claim a position is narrower than it was and the clocks are running at different speeds depending on where you sit. Model providers will ship the next capability jump very quickly. And meanwhile, enterprise procurement will often lock in platform decisions that run for 6 to 18 months and hyperscaler capex is running on a different clock yet again needing to show returns by two or three years out. If you're in the middleware layer and you're looking at the contract lockins, you're looking at the hyperscaler capex, I just want to comfort you. The clock that matters is the first one. The one where we talk about capability jumps. Every model generation that collapses your differentiation is going to narrow the window for repositioning. Everybody else is renting their space and the landlords are getting hungrier. So, you have to be super super savvy about where you position yourself and you need to recognize that the clock is ticking on a generic middleware position. If you are working at a company that is in this middleware business and you're listening to this video and you're thinking, "Oh man, I'm not a CEO, but I'm in trouble." You're probably right and you should probably take proactive action now and not just sit there and wait. If you are an engineer or a product person or a builder, I've given you a lot of good ideas to go dig into this. If you're a solo operator, if you're a founder, this is your roadmap to start thinking about strategically how you operate in a world with hungry hyperscalers. And if you're in seauite or leadership, this is something you need to be putting into board conversations as a risk. It's irresponsible not to. In a world where the hyperscalers need downstream tokens, the rest of us have to figure out how to hunt for tokens that are going to be ours. Good luck with that.