Perplexity: The AI Prize Is the Switch, Not the Model
Perplexity never built a frontier model; it built the switch that decides which model, and now which machine, does each step of your work.
Executive summary
The most important thing Perplexity has built in 2026 is not a model or a search engine. It is a switch: software that breaks a task into steps and decides, step by step, which AI model does the work and, since August, which machine it runs on, a frontier model in a data centre or a small open model on the laptop in front of you. If frontier models are becoming interchangeable parts, the company that decides which part does each job holds the position the model makers are fighting to keep. Perplexity, which has never trained a frontier model, is the clearest working template of that idea, and its revenue run rate has quadrupled this year to $1 billion.
The thesis that Perplexity is building a smart router across cloud and edge compute is correct on direction and early on economics. In under five months its local-cloud router went from a stage demo to three shipping products, and the pieces are concrete: a cloud planner, a privacy gate on the device, open-weight local models and, since November, open decision models built to classify and route. But today the router splits work mainly by privacy and by the user's permission, not by price, and the most sophisticated version of the cost-and-latency optimisation its chief executive describes has not shipped. The economics run through a quieter mechanism: work done on your own hardware costs Perplexity nothing.
Others are building routers too. OpenAI routes between its own models in the cloud; Apple and Google route between the phone and their own data centres; Microsoft gives developers the parts to do both on Windows. Each routes inside a stack it owns. Perplexity is the one routing across rivals: models from competing labs, chips from competing vendors, and hardware the customer already paid for. That neutrality is its edge and its exposure, because the companies that own the operating systems can build the same switch for their own platforms. Where this points is a market in which the model is a supplied part and the switch is the prize; whether a neutral company or the owner of the device ends up holding it is the question of the next two years.
Previously: from answer engine to a computer that does the work
What we covered before, in our first Perplexity episode, number 2, in April, was its turn from a citation-first answer engine into Computer, an agent that does the work, and the open question of whether that makes it an operating system. That episode, Perplexity: From an AI Browser to the Future of the Computer, published on 24 April 2026, made three points worth carrying forward.
- Perplexity began as an answer engine that cited its sources and positioned itself against Google search, and the pressure of competing with Google on search alone pushed it to find a larger role.
- Its move was from retrieving information to executing tasks: Computer, launched on 25 February 2026, takes a goal and works through it, from browsing the web to operating it.
- The open question was whether this was a better search tool or the early shape of a new operating system for knowledge work.
What has changed since is that the operating-system idea now has a specific shape. Computer was always a conductor that handed parts of a job to different models. In 2026 Perplexity has extended that conductor in one decisive direction: it now also decides where the work physically runs. That is the move this episode examines.
So what: the first episode asked whether Perplexity was becoming an operating system; the second answers that the part of an operating system it is building is the scheduler, the piece that decides what runs where.
The model is becoming the cheap part
The case for a router starts with the case against the model as a moat. The episodes where we talk about frontier models becoming commoditised are our AI bubble episode, number 49, our Kimi K3 episode, number 33, and our Apple episode, number 30, and what it means is that value moves from the model to whoever decides which model does each job.
In our AI bubble episode, number 49, from late August, we argued that the exposed layer of the AI market is the pure-model lab: a company whose only product is a model, renting its compute and pricing against open-weight alternatives that cost almost nothing to use. In our Kimi K3 episode, number 33, from July, we followed a Chinese lab that reached the frontier and then released its model's weights for anyone to run. And in our Apple episode, number 30, also from July, we described compute splitting in two: a thin, expensive frontier for the hardest reasoning and a fast-falling commodity tier that runs on hardware people already own.
The independent measurements point the same way. Epoch AI, which tracks model capability, found in May that since January 2026 the most capable open-weight models have trailed the best closed models by an average of about four months on its capabilities index. Its tracking of inference prices shows the cost of reaching a fixed level of capability falling by roughly 40 times a year. A model that is best in class today is, within a season, matched by one anyone can download, and within a year the same capability costs a small fraction of what it did.
None of this means frontier models are worthless. The hardest reasoning still needs the biggest models, and the labs that build them are still the most valuable private companies in the industry. It means a single model is rarely the whole answer to a task, and that most of the steps inside a long task do not need the most expensive one. Perplexity's chief executive, Aravind Srinivas, has built his strategy on that observation. "The orchestration is the product. The model is a tool," he told Fortune on the day after Computer launched, adding that "the model layer is the most competitive it's ever been."
So what: if models are suppliers competing on price and a four-month lead, the durable position is the buyer that chooses among them, and Perplexity has organised itself to be that buyer.
In three months, the router moved onto your machine
What Perplexity is building is a neutral layer that breaks each task into steps and decides which model, and now which machine, does each one. The first half was already true in February: at launch, Computer orchestrated 19 models, using Anthropic's Claude Opus 4.6 for orchestration and coding, Google's Gemini for deep research, xAI's Grok for quick lightweight tasks and OpenAI's GPT-5.2 for long-context recall. Srinivas compared it to hiring: "When you build a team, you don't build a homogenous group where everyone has the same skills."
What has changed over the last three months is that Perplexity moved the router onto the user's machine, shipping Portable Computer in August, hybrid compute on Macs in September, a fully local Computer in October and open decision models in November, while its revenue run rate reached $1 billion.
The sequence started on stage. On 2 June, during Intel's keynote at Computex in Taipei, Srinivas demonstrated what Perplexity called the first "hybrid local-server inference orchestrator": while working through confidential deal documents, local models running on Intel's Core Ultra Series 3 chips decided which information stayed on the laptop and which could go to frontier models in the cloud. The company said the system would be chip-agnostic. The next day he gave CNBC the line that frames the whole strategy: "The data center is coming to your laptop."
Three products followed. On 25 August, Portable Computer put the whole agent, its orchestrator and its models on Nvidia's DGX Spark desktop, with an open model of 27 billion parameters doing the work and the cloud called only when the user approves. On 1 September, hybrid compute came to the Mac: a task can start in the cloud, where a frontier model plans, searches and reasons, and hand the steps involving private files to a model running on the Mac. On 20 October, Perplexity made Computer run fully locally on Apple silicon, with the model, the agent and the execution all on the machine.
Around the router, the agent itself became more persistent. In September Perplexity added Automations, which let Computer pick up recurring tasks on a schedule, and on 4 November it launched Teammates, agents that own ongoing work such as research, bug triage or a marketing campaign, starting with subscribers to its Max plan. Persistent agents matter for routing because they multiply the number of decisions: an agent that works every day for weeks makes thousands of small choices about what to run where, and most of them are the cheap, repetitive kind that suit local hardware.
Two November releases complete the picture. On 12 November Perplexity published Decider, a family of small open-weight "decision models" that return probabilities rather than prose, the kind of fast classifier a router needs at every fork. On 16 November Srinivas told Semafor that Perplexity had crossed a $1 billion annual revenue run rate, up from under $250 million at the start of the year by The Information's reporting, and the company announced Perplexity Frontier Research, a lab with more than 10,000 GPUs whose agenda includes pre-training.
So what: in the space of one season, Perplexity stopped being only a conductor of other companies' cloud models and became a conductor that also decides whether the work leaves your computer at all.
How it is built: a cloud planner, a local gate and a small decider
How Perplexity is building it: a frontier model in the cloud plans the task, a small classifier on the device screens each step for private data, and a local open-weight model handles what must stay on the machine. The design is easiest to see in hybrid compute on the Mac, the most widely available of the three products.
The cloud does the work that needs a frontier model: planning, web search and heavy reasoning. The Mac does the work that should not leave it: reading private files, handling sensitive information and acting inside the user's own apps. Between them sits what Perplexity calls a privacy gate. An on-device classifier flags names, addresses, account numbers and secrets before anything leaves the machine, and the gate then masks the detail, keeps the step local, refuses the action or asks the user for consent. Credentials, card numbers and government identity numbers get the strictest treatment, and enterprise administrators can set rules for a whole organisation and audit when information leaves a device. Perplexity open-sourced that classifier, a model of about 600 million parameters, small enough to run on any machine that can run the agent.
The local models are not Perplexity's own frontier models; they are open-weight models from other labs, sometimes retrained by Perplexity. Hybrid compute launched with Google's Gemma 4 E4B, Alibaba's Qwen3.6 35B-A3B and a Perplexity model post-trained for Computer. The hardware bar is real: hybrid compute needs an Apple silicon Mac with at least 24 GB of memory, and the fully local mode is recommended for 36 GB or more.
Decider is the newest piece and the most telling. A decision model is a general-purpose classifier: asked a question about some content, it returns a probability for each possible answer instead of writing text. Perplexity says it already runs a multi-head classifier on every query, with one head deciding whether a query needs a web search and another planning how deep the research should go, and it lists directing requests to smaller or larger models among Decider's uses. On its own tests the larger Decider, built on a 27-billion-parameter Qwen model, averaged 91.6% accuracy across 63 query-classification tasks, 6.54 points better than the classifiers Perplexity runs in production. A smaller version is distilled into a 4-billion-parameter model. Both are released as open weights. In our Jev episode, number 61, from September, we described a fast, text-free decision model built to sit beside a large language model; Perplexity benchmarks Decider directly against Jev, which shows that the router's brain is becoming a product category of its own.
The pattern across all of it is the same: Perplexity builds the small, fast parts that decide, and buys or borrows the large parts that think. Its competitive assets are the harness, the classifiers and the integration with the user's machine, not a frontier model.
So what: the router is not a slogan but a stack of shipping components, and the component that matters most, the gate that decides what leaves the device, is one Perplexity has chosen to give away, which tells you it is betting on the system around it rather than the part itself.
The thesis holds on direction, not yet on price
The thesis is correct on direction, because Perplexity is building for routing compute between edge and cloud, but today its router splits work mainly by privacy and the user's permission, not yet by cost. The distinction matters for whether Perplexity has found the winning model or only an early version of it.
Start with what Srinivas says the router is for. In the CNBC interview he named one metric that will decide the winner: whoever delivers the "most token value per watt per user." Winning, he said, means "balancing accuracy, latency, cost, privacy and intelligence all together," and "this is an orchestration problem." He described Perplexity's product as a "hybrid neutral orchestration layer" and "the most versatile operating system," one that works across models, chips and hardware. That is the full thesis in his own words: an operating system whose core job is to schedule intelligence across machines on five criteria at once.
The operating-system language is more precise than it sounds. In a conventional operating system, the scheduler decides which program runs on which processor core and when; the user never sees it, and nothing works without it. Perplexity's router does the same job one level up. The programs are the steps of a task, the processors are models and machines, and the rules weigh accuracy, latency, cost and privacy. An app that owns that decision is as close to an operating system as an app can get without owning the device.
Now look at what ships. In hybrid compute the user opens the model picker, selects hybrid, and chooses both the local and the cloud model for the task. The classifier decides which steps are private. Portable Computer and the fully local Mac mode run locally by default and reach the cloud only when the user authorises it. All three products are useful, and all three route on one criterion, sensitivity of data, with the user setting the rest. The automatic weighing of cost and latency against accuracy that Srinivas describes is the direction of travel, not yet the shipping behaviour.
The economics, meanwhile, run through a mechanism that is easy to miss. Perplexity says plainly that on-device work "doesn't consume credits" and that local-first means "people no longer have to ration what they use intelligence for." Every step that runs on the user's Mac or DGX Spark is a step Perplexity does not pay a cloud provider for. A router that moves work to hardware the customer already owns lowers the company's cost of serving a subscriber, which matters for a business that buys most of its intelligence from other labs. Srinivas acknowledged that dependence in June, attributing the tripling of Perplexity's revenue run rate in the first five months of the year to "model advances that have been made by Anthropic."
Our earlier episode on this topic, our Apple episode, number 30, argued that compute was splitting into a costly cloud frontier and a nearly free tier on your own hardware, and Perplexity's products since then substantiate it by building the switch between the two. That July episode concluded that most production queries never need a frontier model and that the real obstacle to running locally was the harness, the scaffolding around the model, rather than the model. Perplexity's 2026 releases are close to a direct test of both claims: it built the harness, shipped it on consumer machines, and kept the frontier in the cloud for the steps that need it.
There is one signal pulling the other way. Perplexity Frontier Research, announced alongside the $1 billion milestone, lists pre-training among its pillars and promises to release models. A company fully convinced that models are interchangeable parts would not need 10,000 GPUs of its own. The more likely reading is insurance: a router that depends entirely on rivals' models is exposed if those rivals raise prices or restrict access, and owning some capability, especially in the small models that run on the device, reduces that risk.
So what: the thesis that the winning model is a smart router across cloud and edge is supported by what Perplexity has built, but the version that would make it decisive, automatic routing on cost and speed as well as privacy, is still a promise, and the company is quietly hedging on models while it waits.
Everyone is building a router; almost nobody a neutral one
Others are doing it too: OpenAI, Apple, Google and Microsoft all route work between models, but each routes within its own stack, while Perplexity routes across rival labs' models and the user's own hardware.
OpenAI made routing the default experience of ChatGPT with GPT-5 in August 2025. Its system card describes "a unified system with a smart and fast model that answers most questions, a deeper reasoning model for harder problems, and a real-time router that quickly decides which model to use based on conversation type, complexity, tool needs, and explicit intent." Every model it chooses between is OpenAI's, and every one runs in OpenAI's cloud.
Apple routes across the device line. Its rebuilt Siri, Tim Cook told investors in January, will "continue to run on the device and run in Private Cloud Compute," Apple's own servers, and the personalised Siri is "a collaboration with Google," whose Gemini models underpin the new foundation models. Apple decides what stays on the iPhone and what goes to its servers, but the models and the servers belong to Apple and one chosen partner.
Google offers developers the same split on Android. In April it announced an experimental hybrid inference API in Firebase AI Logic that routes requests between Gemini Nano on the device and Gemini models in the cloud; the first version uses simple rules, preferring the device and falling back to the cloud or the reverse, and Google says smarter routing is planned. Microsoft's Foundry on Windows gives developers local runtimes for built-in, open-source and custom models alongside connections to cloud services, and Qualcomm announced with Hugging Face in June a planned agent that chooses whether a model runs on a Snapdragon device or in a data centre based on performance, cost, privacy and latency.
Placed side by side, the pattern is plain. The platform owners route, but each routes inside a stack it controls: its own models, its own data centres, its own operating system. Perplexity is the one routing across rivals, picking among more than 15 frontier models from competing labs in the cloud and open models from others on the device, and running on Apple and Nvidia hardware today, with Intel laptops shown on stage. Samsung has given it a place on the phone too: the Galaxy S26 ships with a "Hey Plex" wake word alongside Google's and Samsung's own assistants, and Perplexity says it powers search and reasoning for Bixby.
Neutrality is the edge. A company that routes across every lab can always pick the best or cheapest model for a step, while a platform owner has reasons to prefer its own. Neutrality is also the exposure. Perplexity's router lives inside operating systems that belong to Apple, Google and Microsoft, and each of them can build its own switch at the system level, with deeper access to files, apps and hardware than any app can have. Apple has already done so for Siri.
So what: the market is converging on routing; what Perplexity has that the others lack is not the router but the willingness to route to anyone, and that is a position the platform owners can contest but not easily copy without routing to their own rivals.
What could break the template
Three risks sit under the thesis, and none of them is hypothetical.
The first is the router itself. A switch that decides what leaves the device has to be right nearly every time. Reporting on the Computex launch called classification the risk Perplexity must manage: a wrong call could send sensitive material off the machine, or hand a hard task to a local model too weak to do it well. Perplexity's answer is layered: the open classifier, a gate that can mask or refuse, and a request for consent. Publishing the classifier lets outsiders test it, but the company has not published real-world error rates for the routing decision itself.
The second is the hardware bar. Hybrid compute needs a Mac with 24 GB of memory, and Portable Computer runs on a DGX Spark with 128 GB. Most laptops and phones in use fall short, so for now the local half of the router serves a well-equipped minority of professionals, who are also the people most likely to handle confidential files. Perplexity is betting that the bar falls on its own; its Portable Computer post put it as "every chip cycle and every model release pushes this further."
The third is legal and commercial dependence. Perplexity has fought a run of suits from publishers and retailers; WinBuzzer counted nine organisations with active suits against it at the end of May. It won an important round in August, when the Ninth Circuit vacated an injunction Amazon had won against its Comet browser agent, holding that the user, not the agent's maker, is the one accessing the site; the case continues in the district court. And every frontier step Perplexity routes is bought from a lab that competes with it for the same customers.
So what: the template works in shipping products for a professional niche; making it the default way people use AI depends on routing accuracy nobody has yet measured in public, hardware that is still arriving, and suppliers who are also rivals.
Where this points: the switch becomes the prize
Where it points us is to the routing switch as the new control point of AI: the model becomes a part bought from many labs, the switch that assigns the work becomes the prize, and on today's evidence the advantage goes to whoever owns the device, unless routing stays an open layer that any assistant can plug into.
For the economy, politics and the race between the labs, it means compute cost shifts onto hardware people already own, private and sovereign work can stay local, and the labs compete as suppliers to whoever runs the switch. Each part is worth spelling out.
The economics change first. The industry's record data-centre spending assumes that intelligence is consumed in the cloud. Every task a router moves to a laptop or a desktop box is demand that never reaches a data centre, paid for instead by the person who bought the machine. That does not make cloud capacity redundant, since the frontier steps still need it, but it caps how much of the work the cloud must carry and it rewards chipmakers that sell into personal computers. Nvidia, which sells the DGX Spark that Portable Computer runs on, was reported in August by The Information to be in talks to invest in Perplexity at a valuation above $30 billion, up from the $20 billion of its last closed round; no deal has been announced.
The politics follow. A Perplexity spokesperson put it directly at Computex: "Sensitive and sovereign work can stay local, which changes the need for massive country-level infrastructure." Governments building national AI capacity for reasons of data sovereignty have assumed they need national data centres. A router that keeps sensitive steps on local machines and sends only safe ones to the cloud offers a cheaper route to the same goal, at least for some workloads. It also hands regulators a concrete object to inspect: the rule that decides what leaves the device.
The race between the labs is the third effect. If the router picks the model for each step, a lab's customer is increasingly the router, not the person. Labs then compete on price and on narrow strengths, coding here, long context there, which is exactly the commodity dynamic our bubble episode described. The labs know this, which is why OpenAI, Google and Apple each prefer a router that keeps the choice inside their own stack.
The signals to watch are the ones that decide that contest. The first is whether Perplexity's router starts to choose on cost and speed without the user picking models, which would make the "token value per watt" thesis real. The second and third are whether the operating-system owners open system-level routing to rival models or keep it closed while shipping their own. The fourth is whether open-weight models keep closing on the frontier, which makes more of every task cheap enough to run locally and makes a neutral router more valuable.
So what: the template is visible now, and Perplexity is the clearest example of it, but the template does not guarantee the template-maker the win; the switch will be worth most to whoever controls the place it is installed.
Bottom line
Perplexity is the clearest working example of the idea that, as models become parts, the value moves to whoever decides which part does each job. In 2026 it extended that idea from choosing among cloud models to choosing between the cloud and the user's own machine, and it shipped three products and a set of open routing components in under five months while its revenue run rate quadrupled.
The thesis that a smart router across cloud and edge is the winning model is right about the direction of the industry. It is early on the mechanism: today's router decides mainly on privacy, with the user making the remaining choices, and the cost-and-speed optimisation its chief executive describes is still to come. It is also contested on ownership: Apple, Google, Microsoft and OpenAI are all building routers, each inside its own stack.
That leaves one question that will decide how much the template is worth to the company that drew it. If routing stays a neutral layer that any device can run and any lab can supply, Perplexity is well placed to be its default. If routing becomes a feature of the operating system, the switch will belong to the platform owners, and Perplexity will have shown the industry the way without keeping the prize.
Sources
- Perplexity: Introducing Hybrid Compute on Mac (1 Sep 2026) — the cloud/local split, privacy gate, launch models and hardware requirements
- Perplexity: Introducing Portable Computer for local-first AI (25 Aug 2026) — local-first agent on Nvidia DGX Spark, credits, cloud escalation to 15+ models
- Perplexity: Perplexity Decider (12 Nov 2026) — decision models, internal routing uses, benchmark results, open weights
- Perplexity: Introducing Perplexity Frontier Research (16 Nov 2026) — the research lab, its pillars and GPU count
- TestingCatalog: Perplexity Computer now runs fully locally on Mac (Oct 2026) — the fully local mode and memory guidance
- Unite.ai: Perplexity launches hybrid compute on Mac with local privacy gate (Sep 2026) — the open-sourced PII classifier and its size
- VentureBeat: Perplexity unveils hybrid local-cloud inference system at Computex (Jun 2026) — the Computex demo and the sovereignty quote
- CNBC: Perplexity CEO says one metric will determine who wins the AI race (3 Jun 2026) — token value per watt per user, the neutral orchestration layer, revenue attributed to Anthropic's models
- Fortune: Perplexity CEO explains Computer (26 Feb 2026) — 19 orchestrated models and "the orchestration is the product"
- Latent Space: Perplexity hits a $1B revenue run rate (Nov 2026) — the run-rate milestone and The Information's earlier figures
- SiliconANGLE: Nvidia reportedly eyes investment in Perplexity at a $30B valuation (24 Aug 2026) — the reported talks and the prior $20B round
- Epoch AI: Open models lag state-of-the-art closed models by four months (29 May 2026) — the open-closed capability gap
- Epoch AI: LLM inference price trends — the falling cost of a fixed level of capability
- OpenAI GPT-5 System Card (arXiv) — the fast model, reasoning model and real-time router
- 9to5Mac: Apple confirms Gemini-powered Siri will use Private Cloud Compute (29 Jan 2026) — Tim Cook on device and cloud routing
- Android Developers Blog: Hybrid inference and new AI models are coming to Android (Apr 2026) — Google's rule-based device/cloud routing API
- Microsoft Learn: What is Microsoft Foundry on Windows — local and cloud model execution for Windows developers
- Qualcomm and Hugging Face expand relationship (Jun 2026) — a planned agent that chooses device or cloud
- WinBuzzer: Perplexity tests AI PC privacy with local-cloud router (Jun 2026) — the misclassification risk and the count of active suits
- Jones Day: Ninth Circuit vacates CFAA injunction against Perplexity's Comet AI agent (Sep 2026) — the Amazon ruling of 4 August 2026
- IT Brief: Perplexity launches Teammates AI agents for Max users (Nov 2026) — persistent agents for ongoing work
- 9to5Google: Samsung Galaxy S26 Perplexity integration (Feb 2026) — the "Hey Plex" hotword on Samsung phones
Transcript
Alex: Last Thursday my board asked which AI model we'd picked. One word, they said. I said "all of them," and nobody loved it.
Sam: Which is the right answer, and also the answer that makes a board nervous.
Alex: Right. And on the way home I realised the model isn't the interesting bit. Whoever decides which model gets which job is.
Sam: Hm. I'd have said the opposite in that room. The model is what everybody argues about.
Alex: Today's story says that argument is about to look quaint. The company is Perplexity, which has never trained a frontier model and whose revenue run rate has quadrupled this year. Perplexity: the AI prize is the switch, not the model. Before we get into it, one thing from my week. My head of payments forwarded me a headline on Monday: OpenAI's new wallet taps into the card networks. That's all I've got, a headline, no details yet. And my first reaction was, great, now an AI can buy things from my stores, and my store has no idea who it's talking to.
Sam: Because to you it's a security problem.
Alex: Everything's a security problem to me. My head of security says I've started to sound like her.
Sam: I'd pay to hear that conversation.
Alex: She did not laugh. What about you?
Sam: Mine is the same headline from the other side. Our customers mostly buy the same bag of dog food every month. If a wallet inside a chatbot starts doing that reordering for them, who owns the relationship? Us, or the chatbot?
Alex: And that's today's episode in miniature. Who sits in the middle?
Sam: Then let's go sit in the middle.
Alex: Welcome to Dan's AI Intel, the show where we try to make sense of the fastest, most consequential shift any of us is likely to live through. I'm Alex, with Sam, and today is Perplexity again. If you like the show, follow it wherever you listen, so you don't miss the next one.
Sam: And we've been here before. Perplexity was our episode two, back in April, and a lot has moved.
Alex: Here's the map. Where we left it. Why the model is turning into the cheap part. How Perplexity moved its decision-making onto your own computer in a few months. How that works, whether the economics hold up, who else is building the same thing, and what could break it.
Sam: And the question under all of it is a simple one. Who ends up holding the switch?
Alex: Let me stake my claim, because I think I know. I think the owner of the device wins. Whoever owns the operating system sits closest to your files, your apps, your chip. I don't love saying it, because open usually beats closed when it works. But I said a long time ago that the company owning the whole stack has an advantage nobody beats, and I haven't seen a reason to move.
Sam: And I'll end up on the other side. The neutral layer wins, because it's the only one sitting on my side of the table. Every time I've been locked into one vendor, the price went up the following year. My renewals are the proof. When I asked one vendor for a cap on usage, they told me it was usage-based, which is a polite way of saying no. So hand me the thing that can switch suppliers when mine misbehaves.
Alex: You're saying it because lock-in is your nightmare.
Sam: I'm saying it because I sign the budget. Different thing.
Alex: We'll see where we both end up.
Sam: So, quick recap for people who missed episode two, which came out in April. Alex, you do it, because I'll mix something up.
Alex: You will. So, three things. Perplexity started as an answer engine. It cited its sources and it set itself up against Google search. Second, the pressure of fighting Google on search alone pushed it toward something bigger. It went from finding information to doing the work. That's Computer, which launched in February: you hand it a goal and it works through it, including operating the web for you.
Sam: What I took from April was the shift from answers to actions. I stopped being impressed by answers around the time I'd rather it just did the thing, and then I stopped being impressed by that the first time it did the wrong thing.
Alex: That's the right scar to have.
Sam: And the open question at the end was whether that made it a better search tool or the early shape of a new operating system.
Alex: Yes. And the operating system idea now has a specific shape. Computer was always a conductor, handing parts of a job to different models. Now it also decides where the work physically runs. In a data centre, or on the laptop in front of you.
Sam: So episode two asked whether Perplexity was becoming an operating system.
Alex: And this one says which part of an operating system it's actually building. The scheduler. The piece that decides what runs where.
Sam: That's the part nobody ever sees.
Alex: Which is what I love about it. Nobody has ever sent a thank-you note to a scheduler, and nothing works without one.
Sam: You're about to go on a tangent, aren't you.
Alex: A short one.
Sam: Alex.
Alex: Fine. I'll save it.
Sam: Start me at the claim I resist. The model is the cheap part. In my world the model is the line item that keeps growing.
Alex: It grows on your invoice. It's falling underneath. Our bubble episode, number 49, from late August, said the exposed layer of the market is the lab whose only product is a model. They rent their compute, and they price against open models that cost almost nothing to run.
Sam: And Kimi K3, number 33, in July.
Alex: The Chinese lab that got to the frontier and then handed its weights to anyone who wanted to run them. And the Apple one, number 30, also July, where we said compute is splitting in two. A thin, expensive frontier for the hardest thinking, and a fast-falling cheap tier that runs on hardware people already own.
Sam: So the story we told in the summer was the market. Today's is a company built to live inside it.
Alex: That's it exactly. Epoch AI, who track capability, found in May that since the start of the year the best open models have trailed the best closed ones by about four months.
Sam: Four months. That's a season. And my contracts don't know what a season is. I signed my team's tools for twelve months in January, and some of that is already overpriced for what it does. If the lead is four months, why am I locked in for twelve?
Alex: And Epoch's tracking of prices says the cost of reaching a fixed level of capability falls by something like forty times a year.
Sam: Forty. Per year. Okay, but that's not what I see. What I see is my vendor telling me the new model costs more.
Alex: The newest one, sure. The same capability a year later costs a sliver.
Sam: So I'm paying a premium to be early.
Alex: You're paying a premium to be first, and your team might not need first. Most steps inside a long task don't need the biggest model. That's the whole insight. A task is thirty small steps and two hard ones.
Sam: And right now I pay the hard-step price for all thirty-two.
Alex: At my company too. Our engineers point the best model at everything because it's the default and nobody wants to be the person who picked the cheap one and got blamed.
Sam: Nobody ever got fired for buying the expensive model.
Alex: That's the sentence my CFO is going to quote at me in January. Now, none of this says frontier models are worthless. The hardest reasoning still needs the biggest models, and the labs that build them are still the highest-valued private companies in the industry. The claim is smaller. A single model is rarely the whole answer to a task.
Sam: And Perplexity's boss, Aravind Srinivas, has built everything on that.
Alex: He told Fortune, the day after Computer launched, "The orchestration is the product. The model is a tool." And he added that the model layer is the most competitive it's ever been.
Sam: A man selling a product that depends on other people's models saying models are tools. Convenient.
Alex: Convenient, and consistent with what he built. At launch Computer orchestrated nineteen models. Claude Opus 4.6 from Anthropic did the orchestrating and the coding, Google's Gemini the deep research, Grok from xAI the quick jobs, and GPT-5.2 from OpenAI the long-context recall.
Sam: Four labs that compete with each other, doing jobs for one company.
Alex: He compared it to hiring. "When you build a team, you don't build a homogenous group where everyone has the same skills."
Sam: That one I buy, though there's a catch. When I hire a specialist, I also get their judgement about when to stay out of a project. A router picks a model for a step. It doesn't know whether the step was worth taking. Somebody still decides that, and she's on my payroll. Also, my creative group would be a disaster if everyone wrote like me.
Alex: Your copywriters would be thrilled to hear that.
Sam: They'd be less thrilled by how much I edit.
Alex: So here's where I want to plant a flag, and I know it's where we'll disagree. If the models are suppliers competing on price and a four-month lead, the durable seat is the buyer who chooses among them.
Sam: Which is exactly my side. The neutral buyer.
Alex: Or the buyer who owns the building the suppliers come to.
Sam: That's a very Alex way of saying "me."
Alex: I deserved that.
Sam: So tell me how it moved from a stage to something on a desk. I like things I can buy.
Alex: This is the part that got me excited. It starts on a stage in June. During the Computex show in Taipei, in Intel's keynote, Srinivas demonstrated what Perplexity called the first hybrid local-server inference orchestrator.
Sam: In English, please.
Alex: A program working through confidential deal documents, with small models running on Intel laptop chips deciding what stays on the machine and what's safe to send to the big models in the cloud.
Sam: So the laptop is the bouncer.
Alex: The laptop is the bouncer. And the next day he told CNBC, and this is the line that frames it all, "The data center is coming to your laptop."
Sam: Which is exactly what a company that doesn't own a data centre would say.
Alex: Careful. We haven't got to that part yet.
Sam: So at Computer, sorry, at Computex, could I buy that?
Alex: No. It was a demo, on Intel laptops, not for sale. And you did the Computer-Computex thing, which is your thing.
Sam: It's a very similar word. Somebody should be fired.
Alex: The product people, or the people who named the show?
Sam: Both. Go on. What came after?
Alex: Three products, in order. On the twenty-fifth of August, Portable Computer. The whole agent, its orchestrator and its models, sitting on a Nvidia DGX Spark desktop. An open model with twenty-seven billion parameters does the work, and the cloud only gets called when the user says yes.
Sam: So that's the first one that's real.
Alex: And then on the first of September, hybrid compute came to the Mac. A task can start in the cloud, where a frontier model plans and searches and reasons, and then hand the private-file steps to a model running on the Mac.
Sam: That one I could actually have.
Alex: If you've got an Apple silicon Mac with enough memory, yes. And on the twentieth of October, the fully local mode. The model, the agent and the execution all on the machine.
Sam: Three products in under five months. What I notice is that every one is sold on keeping things private. Fear sells first. Savings come later, once people trust it.
Alex: Fear sells first is right, and it's my CISO's whole worry. Privacy was the easiest thing to prove, so it shipped first. Cost is the harder promise.
Sam: Hm. That's a big claim from a man who builds things for a living and loves a roadmap.
Alex: Around the router, the agent itself got more persistent. In September they added Automations, so Computer picks up recurring tasks on a schedule. And on the fourth of November, Teammates. Agents that own ongoing work, research, bug triage, a marketing campaign. It starts with people on the top plan, Max.
Sam: A marketing campaign. Owned by an agent. My Director of Lifecycle runs our autoship messaging, and the first thing she'll ask isn't whether it's clever. She'll ask who's on the hook when it sends a thousand wrong emails to people who already cancelled.
Alex: Owned by one. Now, why does that matter for routing? Because a persistent agent multiplies the decisions. One that works every day for weeks makes thousands of small choices about what runs where, and most of those are cheap, repetitive ones that suit local hardware.
Sam: So the router matters more the longer the agent lives. And that's what I'd budget for. Not the single clever task. The thousand dull ones I forgot about.
Alex: Then I'd route my store-systems chatter locally first, the inventory checks and order lookups nobody watches. That's where a bill hides, and the data stays inside.
Sam: And two more pieces landed this week.
Alex: This week, yes. Twelve November, Decider, which we'll come to. And on the sixteenth Srinivas told Semafor that Perplexity had crossed a billion dollars in annual revenue run rate.
Sam: A billion. That's a company that quadrupled in a year.
Alex: It was under two hundred and fifty million when the year began, going by The Information's reporting. And the same day they announced a research lab, Perplexity Frontier Research, with more than ten thousand GPUs. Hold on to that one. It comes back.
Sam: You're teasing me with a loop.
Alex: I'm keeping you listening.
Sam: It's working, annoyingly. I'm already waiting for it. Right, so now you get to be nerdy. How does a task actually move?
Alex: Let's take hybrid compute on the Mac, because it's the version most people could use. Picture two rooms. In the cloud room, a frontier model does what needs a frontier model: it plans the task, searches the web, does the heavy reasoning. In the Mac room, a smaller model does what shouldn't leave the machine. Reading your private files, handling sensitive details, acting inside your own apps.
Sam: And in the doorway between the rooms?
Alex: A gate. Perplexity calls it a privacy gate. A classifier on the device looks at each step before anything leaves, and it flags names, addresses, account numbers and secrets. Then it does one of four things. It masks the detail, keeps the step local, refuses the action, or asks you for consent.
Sam: So the little model is basically a bouncer holding a clipboard.
Alex: With a clipboard and a rule book. Credentials, card numbers and government ID numbers get the strictest treatment. And enterprise administrators can set rules for a whole organisation and audit when information leaves a device.
Sam: Audit. Now that I like. That's a word my finance team understands. But four options is a menu, not a test. My customer service inbox is the test: a name, an address and half a card number in one angry message at eleven at night. Does it mask the card and keep the anger? Because the anger is the part I need to read.
Alex: And my CISO. She stopped me in the corridor this week and asked what that gate would flag before any agent touches our store customer files. I said I didn't know. She said, find out, and don't tell me it's open source as though that's an answer.
Sam: Is it open source?
Alex: The classifier is. A model of about six hundred million parameters, small enough to run on any machine that can run the agent. Perplexity gave it away.
Sam: Why would a company give away the gate?
Alex: That's the bit I'd underline. The decision point for what leaves the device is exactly what they open-sourced. You only do that if you think the money is in the system around it, not the part itself.
Sam: Or you do it so outsiders can test it and your lawyers can say they did.
Alex: Both can be true.
Sam: Mine usually are. And the model on the Mac, whose is it? Please say it's theirs, so I know whose neck my files are in.
Alex: Not theirs. The local models are open-weight models from other labs, sometimes retrained by Perplexity. Hybrid compute launched with Google's Gemma 4 E4B, Alibaba's Qwen3.6 35B-A3B, and a Perplexity model post-trained for Computer.
Sam: So, Google on my laptop.
Alex: Gemma. Not Gemini, which stays in the cloud. Same company, different animal.
Sam: Hardware, though. What's the bar?
Alex: Real. Hybrid compute wants an Apple silicon Mac with at least twenty-four gigabytes of memory, and the fully local mode wants more. That's a well-equipped professional's laptop.
Sam: So the people who can use this are the people with the expensive laptops. Which leaves out my copywriters on their older machines, and they're the ones who'd gain most from work that costs nothing.
Alex: Which are also, as it happens, the people most likely to be handling confidential files.
Sam: Hm. My son has a gaming laptop with a model on it right now.
Alex: Theo has a model on his laptop?
Sam: A small one, for his Minecraft mods. No account, no cloud, and he's very smug about it. He told me the data never leaves the machine, and I realised that's roughly the pitch I'm being sold for a premium.
Alex: A fourteen-year-old who invented the privacy gate.
Sam: He didn't invent the gate. He just doesn't have anything private to protect.
Alex: That's the most accurate review of a privacy feature I've ever heard. Decider is the newest piece, and I think it's the revealing one. It's a family of small open-weight models, published twelve November, and they don't write text. They return probabilities.
Sam: Meaning what? Give me the version I could explain at a stand-up without a diagram.
Alex: You ask it a question about some content, and instead of an essay it hands back a score for each possible answer. Is this a search question, yes or no? How deep should the research go? Should this step go to a smaller model or a bigger one?
Sam: So it's a sorting hat, and a quick one. Which means everything depends on the sort being right, and I've seen what a wrong sort costs. Our own targeting once filed a lapsed customer as loyal, and we sent her a thank-you for her loyalty the week she cancelled.
Alex: A router's brain. Perplexity says it already runs a classifier like this on every query. One head decides whether a search is needed, another how deep to go. The larger Decider, built on a Qwen model, beat the classifiers it runs in production by a bit over six points.
Sam: Slow down. I can feel my eyes doing the thing they do in budget reviews.
Alex: Fair. It beats what they use now, and both sizes are free to download.
Sam: And this connects to Jev, doesn't it. The fast one from that September episode.
Alex: Number sixty-one, the fast text-free one. Perplexity benchmarks Decider against Jev directly.
Sam: So the router's brain is turning into its own category. That surprises me. My martech vendors will start selling that brain as a separate line item, and I can already see the invoice I'll have to defend.
Alex: It's turning into a product. Across everything we've covered, Perplexity builds the small fast parts that decide, and buys or borrows the large parts that think.
Sam: A company whose edge is the harness and the classifiers, not a frontier model. That's a strange thing to build a billion-dollar run rate on, and it's the kind of dependence I flag in my own agency contracts.
Alex: The harness. My word of the month. And here's what it means for me on Monday. I should own the piece of my stack that decides, and rent the piece that thinks.
Sam: Which sounds good until you ask who's responsible when the decider is wrong.
Alex: So here's the thesis, and then I want you to poke it. The idea that the winning model is a smart router across cloud and edge is right on direction. Today it routes mainly on privacy and on what the user permits. It doesn't yet route on cost.
Sam: Say that again, slowly, because that's the sentence I'd be paying for. If it routes on privacy and not on price, then I'm still the one doing the weighing.
Alex: Start with what Srinivas says the router is for. He told CNBC one metric will decide the winner: whoever delivers the most token value per watt per user. Winning means, in his words, balancing accuracy, latency, cost, privacy and intelligence all together, and "this is an orchestration problem."
Sam: Five things at once. That's a lot to ask of anything. I've watched my own team fail at two. Accuracy and cost alone start a fight in every planning meeting I run.
Alex: And he called the product a hybrid neutral orchestration layer, and the most versatile operating system.
Sam: And that's the scheduler line you kept stopping yourself saying, so go on, you've clearly been dying to.
Alex: Yes! Okay, I'm allowed now. In a normal operating system, the scheduler decides which program runs on which processor core and when. You never see it and nothing works without it. Perplexity's router does the same job one level up. The programs are the steps of a task, the processors are models and machines, and the rules weigh those five things. An app that owns that decision is as close to an operating system as an app can get without owning the device.
Sam: Is that the tangent? I braced for twenty minutes and that was under one.
Alex: That was the tangent.
Sam: It was good, and I hate that it was good, because now you'll do it again.
Alex: Thank you.
Sam: And now I'd like to be unfair to it, because that's my job on this show. What ships today?
Alex: In hybrid compute, you open the model picker, choose hybrid, and pick both the local and the cloud model yourself. The classifier decides which steps are private. The other two run locally and reach the cloud only when you authorise it.
Sam: So you pick the models. Then who's doing the routing, exactly, if I'm the one choosing?
Alex: You pick the models. All three products route on one thing, how sensitive the data is, and you set the rest. The automatic weighing of cost and speed against accuracy is the direction, not the behaviour you can buy today.
Sam: So it's a smart router that asks you what to do. I hired that person once. It didn't end well.
Alex: Today it's a very good bouncer and a hopeful scheduler.
Sam: So I wait for the version that actually does the weighing.
Alex: Ship it. That's what I tell my own team, and Perplexity should hear it too.
Sam: Okay, but who pays for that?
Alex: Quietly, Perplexity does less than you'd think. Here's the mechanism people miss. They say plainly that on-device work doesn't consume credits, and that local-first means people no longer have to ration what they use intelligence for. Every step that runs on your Mac or your DGX Spark is a step Perplexity doesn't pay a cloud provider for.
Sam: So when it moves my work onto my machine, my machine pays for it. And my electricity. Which is fine for one customer, and a different conversation when you're buying machines for a whole team.
Alex: Your hardware, your electricity, their margin. A router that moves work onto machines customers already own lowers the cost of serving a subscriber. It matters because they buy most of their intelligence from other labs. And Srinivas more or less said so in June, when he attributed the tripling of their revenue run rate in the first five months of the year to the model advances made by Anthropic.
Sam: So their growth leans on the labs they route to. Which is a familiar position. My agencies grow when my budget grows, and they're lovely to me right up until the renewal.
Alex: Which is exactly the exposure.
Sam: This is the bit that's landed on me this week, actually. My friend Lauren, who runs a six-chair hair salon, got a note that her booking software is moving to credits. She called me and asked what a credit is, and I said, honestly, I don't know, and neither does the vendor yet.
Alex: And you assumed the worst.
Sam: It's my area of expertise.
Alex: Credits are tokens in a wig.
Sam: That's going straight to her.
Alex: But look at what Perplexity's doing about the other side of it. "Doesn't consume credits" is a promise to Lauren-shaped customers. It says: the local step is free to you.
Sam: And I'd believe that more if I'd seen the router choose it for me.
Alex: Which is the same thing you said about the price. One more thing, and it's the loop I opened. There's a signal pulling the other way. Perplexity Frontier Research, announced alongside the billion-dollar milestone, lists pre-training among its pillars and promises to release models. A company that really thought models were interchangeable parts wouldn't want ten thousand GPUs of its own.
Sam: If a vendor told me the model's a tool and then built one, I'd ask what else they're hedging on, because I don't renew on trust.
Alex: I'd read it as insurance. A router that depends entirely on rivals' models gets squeezed if those rivals raise prices or restrict access. Owning some capability, especially the small models that run on your device, reduces that risk.
Sam: Hedging.
Alex: It's what I do with my own vendors. I don't want one contract I can't walk away from. And our Apple episode, number thirty, argued compute was splitting into a costly cloud frontier and a nearly free tier on your own hardware. That July episode said most production queries never need a frontier model, and that the obstacle to running locally was the harness, not the model.
Sam: And Perplexity's 2026 is close to a direct test of both.
Alex: It built the harness, shipped it on consumer machines, and kept the frontier in the cloud for the steps that need it.
Sam: So we were right.
Alex: Early, and partly right. My board would take that.
Sam: Right, so Perplexity does this. Does anyone else?
Alex: Everyone does, and that's the point. OpenAI made routing the default experience of ChatGPT with GPT-5, in August 2025. Its system card describes a fast model, a deeper reasoning model, and a real-time router that picks one based on the conversation's type, complexity and tool needs.
Sam: And every model it picks between is OpenAI's. Hold on, though. A router that sends me to the fast model is making a pricing decision as much as a quality one. Who tells me the cheap model was good enough? I've been burned before by a vendor "optimising" my experience.
Alex: Nobody tells you, and that's the argument for a router that isn't selling you a model. And every one runs in OpenAI's cloud. Apple routes across the device line. Tim Cook told investors in January the rebuilt Siri will keep running on the device and in Private Cloud Compute, Apple's own servers, and the personalised Siri is a collaboration with Google, whose Gemini models sit under the new foundation models.
Sam: So Apple decides what stays on the iPhone and what goes up. And as the person who has to explain where a customer's data went, that's the part I'd want in writing.
Alex: That's fair, and the answer is a written policy, which is what Apple's setup is not. The models and the servers belong to Apple and one chosen partner. Google does it for developers on Android. In April it announced an experimental hybrid inference API in Firebase that routes requests between Gemini Nano on the phone and Gemini in the cloud. The first version uses simple rules. Prefer the device, fall back to the cloud, or the reverse.
Sam: So rules, not a scheduler. Which I'd have bet on, because anyone shipping fast ships the dumb version first, and I respect that more than a clever one that arrives next year.
Alex: Rules, and rules are easy to audit, which my head of security would like. Google says smarter routing is planned. Microsoft's Foundry on Windows gives developers local and cloud models side by side.
Sam: Everyone's building the bouncer. I'd just like to know which of them lets me pick my own club.
Alex: You don't, in any of them. That's the shape. Each one routes inside a stack it owns, with its own models, its own data centres, its own operating system.
Sam: And Perplexity? Because so far you've described four companies doing the same thing in four walled gardens.
Alex: Perplexity's the one routing across rivals. More than fifteen frontier models from competing labs in the cloud, open models on the device, Apple and Nvidia hardware today. And on phones, the Galaxy S26 ships with a "Hey Plex" wake word, next to Google's and Samsung's own assistants, and Perplexity says it powers search and reasoning for Bixby.
Sam: "Hey Plex." I'm not saying it out loud in a supermarket. And three assistants fighting over one wake word is how a customer ends up arguing with a phone. Who wins that, the maker or the user?
Alex: The maker, usually, which is a problem for you and a feature for them. So neutrality is the edge. A company that routes across every lab can always pick the best or cheapest model for a step. A platform owner has reasons to prefer its own.
Sam: That's the piece I'd hold onto. If I'm the buyer, I need a router whose incentive isn't to push me toward its own model. Every vendor I've ever had gets less friendly the moment I can't leave.
Alex: And it's also the exposure. Perplexity's router lives inside operating systems owned by Apple, Google and Microsoft, and each of them can build a switch at the system level, with deeper access to your files, your apps and your hardware than any app will ever have.
Sam: Has anyone actually done that yet?
Alex: Apple already has, for Siri.
Sam: So you've got a guest in somebody else's house who's built the best furniture, and the landlord's started making furniture.
Alex: That is my entire case, thank you.
Sam: It's not your case. It's a metaphor with a case-shaped hole in it.
Alex: It's a good metaphor.
Sam: It's a fine metaphor. Here's mine, though. The landlord can build furniture for the landlord's own rooms. A router that has to route to its own rival has to be good at being a router, and Alex, nobody at Apple is going to route me to Google's rival.
Alex: They partnered with Google for Siri.
Sam: One partner they chose. I'm not asking for one partner. I'm asking whether the switch will send my customers to whichever model is cheapest this month.
Alex: That's the neutral's edge, I give you that. It's a position the platform owners can contest but not easily copy without routing to their own rivals.
Sam: There. We agree on something. Write it down. That's the part I'd put in a vendor brief: tell me who you route to, and tell me who you refuse to route to, and why. A neutral router that quietly favours one lab is just a landlord in a nicer coat.
Alex: I'll put it on the record.
Sam: Don't. Go on, scare me. What could break the template?
Alex: Three things, and none of them is hypothetical. First is the router itself. A switch that decides what leaves the device has to be right nearly every time. Coverage of the Computex launch called classification the risk Perplexity has to manage. A wrong call could send sensitive material off the machine, or hand a hard task to a local model too weak to do it.
Sam: And do they publish how often it's wrong?
Alex: No. Perplexity's answer is layered: the open classifier, a gate that can mask or refuse, and a request for consent. And putting the classifier out lets outsiders test it. But the company hasn't published real-world error rates for the routing decision itself.
Sam: That's the number I'd want in the contract.
Alex: That's the number she'll want before anything goes near customer data, and she's right.
Sam: Second?
Alex: The hardware bar. Twenty-four gigabytes of memory for hybrid compute, a hundred and twenty-eight on the desktop box. Most laptops and phones in use fall short, so for now the local half of the router serves a well-equipped minority of professionals. Perplexity is betting the bar falls on its own. Its Portable Computer post said every chip cycle and every model release pushes this further.
Sam: That bet I'd take. My copywriters' laptops are older than some of my interns, and they've survived everything so far.
Alex: There's a thing there I'd push on, though, because I've lost money on that assumption. Cheaper arrives later than the slide says.
Sam: You lost money on a hardware prediction?
Alex: I lost two years on one, and an apology to my CFO. We can leave it there.
Sam: No, we can't.
Alex: We can. The third risk is legal and commercial. Perplexity has fought a run of suits from publishers and retailers. One count had nine organisations with active suits against it by the end of May. But it won an important round in August, when the Ninth Circuit vacated an injunction Amazon had won against its Comet browser agent. The court held that the user, not the agent's maker, is the party accessing the site.
Sam: Which matters for everybody building an agent, not just them. My brand is going to have agents shopping its site, and the idea that the user is the party accessing it, not the agent's maker, is what keeps my legal team calm and my pricing page honest.
Alex: Hugely. And the case goes on in the district court. The last piece is the layer under everything. Every frontier step Perplexity routes is bought from a lab that competes with it for the same customers.
Sam: So the supplier is also the rival. I know that relationship. It's called my agency contract.
Alex: Which is how you get a company with ten thousand GPUs and a hedge.
Sam: So, what's the summary? That's three risks and I've lost track of the order, and I'd like to take it back to my team.
Alex: One: accuracy nobody's measured in public. Two: hardware that's still arriving. Three: suppliers who are also rivals.
Sam: And for my team, what does that do on Monday? Name one thing I can say in a stand-up.
Alex: It means you try it on the work that can go wrong without hurting anyone. Not the quarterly numbers. The dull stuff.
Sam: The dull stuff my juniors do. Which is, I'd point out, most of what my juniors do.
Alex: That's the part that scares me too.
Sam: My junior copywriters are the dull stuff. I know exactly who does the first draft of an autoship email, and I'd like her to stay employed.
Alex: Better stake than mine.
Sam: Don't undersell it. A budget line is somebody's job.
Alex: So where does it point? My stores run on phones and laptops I can control, so the control point is the device. The model becomes a part I buy from many labs, and whatever assigns the work becomes the prize. Unless routing stays an open layer that any assistant can plug into, the advantage goes to whoever owns the device.
Sam: And that's your case in one sentence. I'll say it's clear. I won't say it's comfortable.
Alex: It's the sentence I'd have written before the show.
Sam: Here's what worries me about it. Where my customers start. If the first thing they open is an assistant on their phone, not our site, then whoever owns that assistant owns my storefront. That's why a neutral layer matters to me. It's my only shot at the assistant not belonging to one company.
Alex: And my answer is that the phone maker is who gets to decide whether that happens.
Sam: Which is the whole fight, in one sentence. So convince me what it means beyond the tech. The economy, the politics, the labs.
Alex: Economics first. The industry's record data-centre spending assumes intelligence is consumed in the cloud. Every task moved to a laptop is demand that never reaches a data centre, and my cloud bill is the thing that would notice. It doesn't make the cloud redundant, since frontier steps still need it, but it caps how much the cloud carries.
Sam: Which is how you end up with Nvidia and Perplexity in the same sentence. And if the work moves onto laptops, then the people who sell laptops and chips quietly win, while the people who sign cloud contracts, like me, get a better negotiating position.
Alex: The Information reported in August that Nvidia, which sells the DGX Spark that Portable Computer runs on, was in talks to invest in Perplexity at a valuation above thirty billion dollars, up from twenty billion in its last closed round. No deal has been announced.
Sam: No deal. Noted.
Alex: And politics. A Perplexity spokesperson said at Computex, "Sensitive and sovereign work can stay local, which changes the need for massive country-level infrastructure."
Sam: That's the claim I'd want tested in public. Every town that's been promised a data centre was told it was needed.
Alex: If sensitive steps stay local, governments might need less of it. For some workloads.
Sam: Some. That word belongs on the brochure.
Alex: And it hands regulators a concrete object to inspect. The rule that decides what leaves the device.
Sam: That's my audit again. And I'd add the unglamorous version. A rule a regulator can read is a rule a company can be held to. A rule that lives in a black box is a promise, and I've signed enough promises from vendors to know the difference.
Alex: That's your audit again, and I'll take it.
Sam: And the labs? If I were one, I'd be furious. You spend a fortune building the thing and the customer never learns your name.
Alex: If the router picks the model for each step, a lab's customer is increasingly the router, not the person. Labs compete on price and on narrow strengths. Coding here, long context there. It's the commodity dynamic from the bubble episode, and it's why OpenAI, Google and Apple each prefer a router that keeps the choice inside their own stack.
Sam: So how do we know who's right? Name things I can look at.
Alex: Four. One, whether Perplexity's router starts choosing on cost and speed without you picking the models. That's what makes token value per watt real. Two and three, whether the operating system owners open system-level routing to rival models or keep it closed. And four, whether open models keep closing on the frontier, which makes a neutral router more valuable.
Sam: The first one's the one I like, because it doesn't need a lawyer or a conference. If a release picks the cheaper model for me and my bill drops, I'll believe in the router. Until then it's a bouncer.
Alex: Next Computer releases. The OS owners' one, next year's developer conferences.
Sam: So we'll know in a year, which is about how long it takes my team to finish one planning cycle. I'd rather know sooner, but I'll take a year with a clear thing to look for.
Alex: More like the next two years.
Sam: That's not a prediction, that's a fortune cookie. Okay. Verdict time. You first, because you've been polite for an hour and it's unnatural.
Alex: I'm still on the owner of the device. I think the switch is worth most to whoever controls the place it's installed. But you moved something. The thing about enterprise administrators setting rules and auditing what leaves a device is the single thing the platform owners haven't shown me. If my CISO gets that from a neutral layer and not from the phone, I'd take it.
Sam: And I moved too. I came in saying the neutral layer wins because it's on my side of the table. I'm still there on price, and on lock-in, and on the fact that I've never once been glad a vendor was the only option. But you've got me on the landlord point. If the landlord decides what the tenant can see, a neutral layer that lives inside somebody else's operating system is only neutral until they say it isn't.
Alex: So we split.
Sam: We split. I'll use the neutral one until my phone tells me to stop.
Alex: And I'll watch whether it's allowed to.
Sam: That's a very you way to end.
Alex: One thing to do on Monday. Find out which of your steps are dull and cheap and private, because that's the work a router will pull off your cloud bill first. The thing I'm taking from this is the question I'll ask every vendor now: who decides which model touches my data?
Sam: And mine is simpler. Whose switch is it, and can I leave it? That's what I'll ask my vendors on Monday. And it's what I'll tell my junior copywriters to ask about the agents they're about to work beside, because the people who do the dull work are the first to notice when the switch is wrong.
Alex: So, the recap. Perplexity never built a frontier model. It built the switch that decides which model, and now which machine, does each step. In under five months it went from a stage demo to three products and a set of open pieces, while its revenue run rate quadrupled.
Sam: It's right that value moves to whoever decides which part does which job. It's early because today's router decides mostly on privacy, and the cost-and-speed part is still a promise. And it's contested because Apple, Google, Microsoft and OpenAI are all building one inside their own walls.
Alex: And the whole thing hangs on one question. If routing stays a neutral layer any device can run and any lab can supply, Perplexity's well placed to be its default. If it becomes a feature of the operating system, the switch belongs to the platform owners, and Perplexity will have shown the industry the way without keeping the prize.
Sam: Thank you so much for listening. We hope you came away seeing a bit more clearly where all of this is heading. It's a fast, complicated picture, and that's what makes it worth following closely.
Alex: And one note on how this is made. It's AI-generated, and so are we. Dan builds a custom stack of AI tools to research, analyse, verify and illustrate the questions worth understanding, mostly to learn them himself, and publishes it for anyone who'd like to follow along. AI-assisted, fact-checked, worth a second look.
Sam: Before we go, one useful thing you can do: follow the show. Whatever app you're listening in, there's a follow or a plus button. One tap, free, and you get each new episode the moment it lands. For a small independent show, a follow is the biggest single thing that helps it reach other people trying to make sense of all this. So if this has been worth your time, go ahead and hit follow.
Alex: And one last thing. If someone in your life keeps asking where AI is heading, sending them this episode is the kindest thing you can do, for them and for the show. A small show lives on every share.
Sam: Until next time.