Signals Inbox·August 25, 2026·AI Knowledge & Search

Why is Nvidia investing in Perplexity?

Nvidia is investing in Perplexity because it is becoming a fast-growing, compute-hungry independent AI platform that can buy Nvidia infrastructure, distribute Nemotron and keep more AI demand outside hyperscalers that own competing chips.

Q1Is Nvidia actually investing in Perplexity again?

Nvidia is currently discussing another investment in Perplexity, and the talks appear serious enough to put a valuation above $30 billion on the table.

The Information's latest reporting says the proposed equity round could be worth billions of dollars and would value Perplexity more than 50% above its previous financing. The deal is still being discussed, so we should treat the valuation and Nvidia's participation as proposed terms rather than a completed transaction.

The relationship itself goes back much further. Nvidia joined Perplexity's $73.6 million Series B in early 2024, when the startup was valued at roughly $520 million. PitchBook data later reported by TechCrunch shows Nvidia participating in most of Perplexity's subsequent rounds, including the financing that valued it at $18 billion in 2025. Nvidia skipped the following $200 million raise that took Perplexity to about $20 billion.

The latest discussions are much more interesting because the commercial relationship has deepened at the same time. Perplexity has joined Nvidia's Nemotron work, plans to use Nvidia's new infrastructure and has become a distribution point for Nvidia models. The Information also reports that the two companies have been meeting several times a week on hardware and software work.

So this latest round would extend a relationship that has moved well beyond Nvidia owning a passive startup stake.

Perplexity financing and Nvidia participation

Q2Why would Nvidia pay a $30 billion-plus valuation for Perplexity?

Nvidia can justify paying much more for Perplexity today because Perplexity's revenue has been rising far faster than its valuation.

The Information reports that Perplexity's annualized revenue has climbed from below $250 million at the beginning of the year to more than $750 million now. That is more than a tripling in well under a year. The proposed valuation, by comparison, would rise a little more than 50% from the roughly $20 billion level of Perplexity's previous financing.

The gap is striking. Investors are being asked to pay substantially more for the company, but Perplexity has grown into the price even faster.

The trajectory becomes clearer when we go back another year. TechCrunch reported in early 2025 that Perplexity had reached roughly $100 million of annual recurring revenue while discussing an $18 billion valuation. The Financial Times later put annualized revenue at around $150 million when the company reached that $18 billion level. By the time Perplexity raised at $20 billion, TechCrunch reported revenue approaching $200 million.

Now the figure is above $750 million.

That still leaves Perplexity with an extremely rich valuation. Using $30 billion and $750 million as the reported floors gives us roughly 40 times annualized revenue. Very few software businesses can support that multiple for long unless growth stays exceptional.

But the valuation discussion looks less absurd once we look at the direction of the multiple. Perplexity's price has continued rising while its revenue multiple has fallen sharply.

For Nvidia, there is an extra layer. Nvidia can make money from Perplexity's growth through the equity stake and through the computing, networking, CPUs, models and software that a larger Perplexity may consume. A normal venture investor gets only the first part.

Q3Is Nvidia really betting on Perplexity Search?

Nvidia's Perplexity bet now reaches far beyond AI search, with Perplexity Computer becoming the more strategically interesting part of the company.

Perplexity originally earned attention by building a conversational search engine before web search became standard inside the largest chatbots. Search still drives a large amount of usage, but the product has been changing quickly.

Perplexity Computer can break a task into subtasks, call different AI models, search the web, run code, use company data and create finished work. Perplexity says Computer now works across more than 20 models, hundreds of connectors and local applications. It has expanded from the web into Microsoft 365, Macs and Windows PCs.

The company has kept pushing in that direction lately. Personal Computer arrived on Windows, giving the agent access to local files, Microsoft 365 and web tools. Projects added persistent files, memory and shared agent workspaces. Computer now also works directly from email, so a user can forward a thread or send a task and receive the result back in the same conversation.

Those releases tell us where Perplexity wants the product to go. The company is trying to sit between a user and a growing collection of models, applications, files and data sources.

That position could generate far more compute per customer than ordinary search. A search answer may involve retrieval and a model response. A Computer task can create several agents, call several models, execute code, browse repeatedly and continue working for much longer.

The Information says Computer has already contributed to Perplexity's recent revenue acceleration, particularly among professionals using it to automate computer-based work.

For Nvidia, that makes the agent side of Perplexity more interesting than the question of whether Perplexity takes another point of conventional search share from Google.

Q4Is Perplexity actually big enough to matter to Nvidia?

Perplexity is still tiny beside Google Search, but its combination of fast growth and compute-heavy usage is already large enough for Nvidia to care.

Perplexity disclosed that it processed more than 500 million queries during all of 2023. By May 2025, CEO Aravind Srinivas said the company was handling around 780 million queries in a single month. That monthly volume was already about 1.6 times the company's entire 2023 total.

Google operates at a completely different scale. Google says it handles more than five trillion searches a year, and its AI Overviews have reached billions of monthly users. Perplexity has a long way to go before conventional search market share becomes the main reason anyone buys Nvidia stock.

But infrastructure suppliers care about the amount of computation generated by a customer, not simply its percentage of the consumer search market.

Perplexity has moved toward workloads that are much heavier than a classic search query. Computer can run long tasks across many models. Model Council, updated again recently, can send the same question to between two and eight AI models and then synthesize the results. Every additional model adds more inference.

Perplexity also says more than 50,000 organizations now use its enterprise products. We should treat company-reported customer counts with the usual caution because "organizations" can include customers of very different sizes, but the number shows how far the company has moved from its original consumer search niche.

Perplexity can therefore become a meaningful Nvidia customer long before it becomes a meaningful threat to Google's overall search volume.

Q5How much computing does Perplexity already need?

Perplexity is already committing hundreds of millions of dollars to cloud computing, which gives Nvidia hard evidence that Perplexity's growth translates into infrastructure spending.

Bloomberg reported that Perplexity signed a three-year, $750 million agreement with Microsoft. Microsoft subsequently confirmed to Reuters that Perplexity had chosen Microsoft Foundry as its primary platform for sourcing models under a multiyear deal.

If the $750 million commitment were spread evenly across three years, the nominal average would be about $250 million a year. That comparison is rough because cloud contracts rarely translate into perfectly even annual spending, but the order of magnitude is useful: $250 million would equal roughly one-third of Perplexity's current reported revenue run rate.

Perplexity also kept its Amazon relationship. A company spokesperson told Bloomberg that spending had not been shifted away from AWS and described AWS as the company's preferred cloud infrastructure provider.

So Perplexity is supporting a Microsoft commitment worth three-quarters of a billion dollars while continuing to use another major cloud provider.

We should also avoid equating those cloud dollars directly with Nvidia GPU purchases. Microsoft now operates Nvidia and AMD accelerators alongside its own Maia chips, while AWS has been pushing Trainium aggressively. Perplexity also buys access to third-party models, so part of the bill represents model usage rather than raw infrastructure.

Even with those caveats, the scale tells us something simple: Perplexity has become a serious buyer of AI compute.

Perplexity infrastructure commitments

Q6Why does Perplexity Computer fit Nvidia's AI agent bet so well?

Perplexity Computer gives Nvidia a real customer for its belief that AI agents will become some of the heaviest users of computing.

Nvidia has designed its newest hardware around that idea. Vera, Nvidia's new CPU architecture, targets work such as agent orchestration, code execution, data processing, tool use and sandbox environments. Nvidia says Vera completes a range of these workloads around 1.8 times faster than x86 systems in its benchmarks.

Perplexity plans to use Vera for agentic workloads, according to The Information.

The fit is unusually clean. Computer creates teams of agents, sends work to different models, runs tools and can keep tasks going in the background. Personal Computer extends that workflow into local files and desktop software. Projects let agents carry memory and files across sessions.

Nvidia CEO Jensen Huang has repeatedly argued that agents will consume enormous amounts of computing because software can work continuously, spawn more tasks and call other software without waiting for a person to issue every command.

This also broadens Nvidia's opportunity beyond GPUs. An agent workflow needs accelerators for model inference, CPUs for orchestration and code execution, networking for moving data between systems, storage for long contexts and software to coordinate the whole stack.

Q7Why does Nvidia want Perplexity to use Nemotron?

Perplexity gives Nvidia something its chip business cannot create on its own: direct distribution for Nvidia's AI models.

Nvidia has become much more aggressive with Nemotron. The company is building open models and the software around them because a healthy open-model ecosystem can stimulate more AI usage without leaving the model layer entirely to OpenAI, Anthropic, Google or Chinese developers.

Perplexity is helping Nvidia do that.

Perplexity joined Nvidia's Nemotron Coalition, where companies contribute data, evaluations and technical work around Nvidia's open models. The Information reports that teams from Nvidia and Perplexity have been meeting several times a week as part of their hardware and software collaboration.

Nemotron has also reached Perplexity users directly. Perplexity's current documentation lists Nvidia's Nemotron 3 Ultra among the advanced models available in Search, alongside models from OpenAI, Anthropic, Google, xAI, Moonshot and others.

That puts an Nvidia model inside a consumer and enterprise product without Nvidia having to build a ChatGPT-scale assistant of its own.

Nvidia's recent behavior shows how seriously it now takes this part of the stack. The company agreed to pay $6 billion to license model-development technology from Poolside and plans to hire more than 100 employees involved with Poolside's Laguna open model. It is also investing $1 billion in Poolside. Separately, Nvidia has discussed investing in Mercor, whose work for Nvidia has grown as Nvidia develops Nemotron.

Perplexity fits the same pattern from another angle. Poolside helps Nvidia build models. Mercor can help provide the data and labor behind them. Perplexity can put those models in front of users.

Together, those deals show Nvidia building more of the machinery around open AI rather than limiting itself to the chips underneath it.

Q8Why does Perplexity using lots of AI models help Nvidia?

Perplexity's multi-model approach helps Nvidia because Nvidia can benefit from rising AI usage even when another company builds the model that wins a particular task.

This is becoming one of Perplexity's defining product choices. Computer routes work across more than 20 models. Model Council lets users deliberately run several leading models against the same problem. Perplexity's current model menu includes systems from OpenAI, Anthropic, Google, xAI, Moonshot, Z.ai and Nvidia.

The latest Model Council documentation makes the philosophy especially clear. Users can choose between two and eight models inside Computer, let each one research independently and then ask Perplexity to synthesize the result.

That architecture is useful to Nvidia because the chip company does not need Nemotron to win every prompt.

Suppose one Computer task uses Claude for analysis, an OpenAI model for orchestration, Gemini for another subtask and Nemotron for a fourth. Perplexity still needs infrastructure throughout the workflow. Some of those models may run on Nvidia hardware directly, some through cloud providers using mixed hardware, and some on competing accelerators. Nvidia can still capture a meaningful share of the overall compute market without controlling the entire model layer.

A vertically integrated company creates a different risk. Google can combine Gemini, Search, Google Cloud and TPUs. Amazon increasingly ties Bedrock and Anthropic workloads to Trainium. Microsoft has its own Maia accelerator.

Q9Are Google TPUs, Amazon Trainium and Microsoft Maia the real reason Nvidia wants Perplexity to succeed?

Google is one of the biggest strategic reasons Nvidia benefits from Perplexity's success, but the issue is broader than Google. Amazon and Microsoft are moving in the same direction: all three combine massive AI distribution with chips of their own.

Google still controls the overwhelming majority of general search. It also owns Gemini, Google Cloud, Android, Chrome and an increasingly mature TPU platform. When Google wins an additional AI workload, it can route more of that workload onto TPUs. When Perplexity wins a workload, the company has to buy infrastructure from outside suppliers and cloud providers.

The scale of the custom-chip push is now hard to dismiss. Anthropic says it currently uses more than one million Amazon Trainium2 chips and has committed more than $100 billion to AWS technologies over the next decade, with up to five gigawatts of new capacity planned. Amazon says Trainium and Graviton already generate a combined annual revenue run rate above $10 billion.

Google has its own huge Anthropic relationship. The AI company plans to use up to one million TPUs and has signed another agreement for multiple gigawatts of future TPU capacity.

Microsoft is moving in the same direction with Maia 200. Microsoft says Maia delivers more than 30% better performance per dollar than the latest hardware previously running in its fleet, and it has already started using the accelerator for inference and internal AI workloads.

All three companies still buy Nvidia hardware. Nvidia remains deeply embedded across AWS, Azure and Google Cloud. But each hyperscaler has a clear financial incentive to move more workloads onto silicon it controls.

Perplexity has no comparable chip program. It can remain a fraction of Google's search size and still matter to Nvidia simply by becoming a large independent buyer of compute.

Q10Why is Nvidia investing in Perplexity, OpenAI, Anthropic and so many other AI companies?

Nvidia is spreading capital across the AI industry because its best outcome is a large ecosystem with many companies spending heavily on compute.

The pace is now hard to miss. PitchBook data reported by TechCrunch shows Nvidia participating in nearly 67 venture deals during 2025, up from 54 the year before, excluding deals made through its separate NVentures arm.

Nvidia's financial statements show the same acceleration at a much larger dollar scale. Its fiscal 2026 annual report records $17.5 billion of purchases of non-marketable equity securities, up from roughly $1.5 billion the previous year. By the end of that fiscal year, Nvidia held more than $22 billion of non-marketable equity investments.

The portfolio covers very different parts of the AI stack. Nvidia has backed model companies, AI applications, data providers, clouds, robotics companies and infrastructure developers.

Perplexity sits in a particularly useful position because it connects several of those layers. It buys compute, distributes models, builds agents and reaches end users.

The breadth of Nvidia's bets also tells us why comparing Perplexity with OpenAI or Anthropic as though Jensen Huang must pick one winner misses the point. Nvidia can benefit from all three growing at the same time.

Q11Is Nvidia just funding customers so they can buy more Nvidia chips?

Nvidia is clearly putting money into companies that can become large compute customers, but the current evidence shows a broader strategy than simply recycling investment dollars into GPU sales.

The circularity concern comes from a real pattern. Nvidia invests in AI companies, those companies need huge amounts of computing, and Nvidia sells the most widely used AI accelerators. Nvidia can therefore help create the demand that later appears in its own revenue.

Perplexity fits part of that loop. Nvidia owns equity in the company, Perplexity plans to use Nvidia infrastructure, and Perplexity's growth creates more compute demand.

Yet several facts complicate the simple version of the story.

Perplexity is spending heavily with Microsoft and continues to use AWS. Its product deliberately supports competing model providers. Microsoft and Amazon can serve some of those workloads using their own chips. We have also seen no disclosed condition saying Nvidia's equity investment in Perplexity has to come back as Nvidia hardware purchases.

Nvidia's newest financing push adds another wrinkle. The Information recently reported that Nvidia has lined up major financial firms including Blackstone, Apollo and Goldman Sachs around a framework that could finance as much as $500 billion of AI infrastructure purchases. The idea is to bring outside capital into AI factories instead of making Nvidia the financier every time a customer needs more capacity.

At the same time, Nvidia is still spending aggressively where it sees technology or ecosystem value. Poolside brings model-development technology and talent. Mercor helps Nvidia's model effort. Perplexity brings users, agents, model distribution and a growing compute bill.

There is circularity in the system. Calling every strategic investment a disguised GPU purchase misses too much of what Nvidia is actually buying.

Q12Why is Perplexity a better Nvidia investment than a random AI app?

Perplexity gives Nvidia an unusually useful mix of revenue growth, compute demand, model distribution, agent workloads and consumer distribution.

Most AI startups offer Nvidia one or two of those things.

Perplexity offers several.

Samsung is a good example. The Galaxy S26 integrates Perplexity as one of the phone's available AI agents. Samsung also uses Perplexity for real-time web search inside the newer Bixby experience. Samsung's own support documentation shows that users can make Perplexity their default digital assistant and invoke it with the phone's side button.

That gives Perplexity distribution outside its own app.

The company is also pushing Computer deeper into professional work. It now operates inside Microsoft 365, works across local Windows files, connects to enterprise tools and can receive jobs directly by email. Perplexity says more than 50,000 organizations use its enterprise offering.

Then there is the technology itself. The Information reports that Nvidia previously considered paying Perplexity billions of dollars to license some of its technology and hire some employees. The exact technology involved has not been disclosed, so we cannot say whether Nvidia was primarily interested in search, retrieval, model routing or agent infrastructure.

The size of the proposed licensing deal still tells us something. Nvidia saw enough technical value inside Perplexity to discuss a transaction measured in billions before the talks shifted toward a more conventional investment.

That combination is rare: Perplexity can be a customer, partner, model distributor, technical collaborator and equity investment at the same time.

Q13Is Perplexity actually worth more than $30 billion to Nvidia?

A valuation above $30 billion remains aggressive, but Perplexity's revenue growth has made the price much easier to defend than its earlier valuation multiples.

At roughly $20 billion, Perplexity was approaching $200 million of annual recurring revenue according to reporting around its 2025 financing. That put the company near 100 times annualized revenue.

The latest figures look very different. The Information now puts annualized revenue above $750 million. Using the reported $30 billion valuation threshold gives a multiple of roughly 40 times revenue before adjusting for the fact that both figures are minimums rather than exact numbers.

Forty times revenue is still expensive. Nvidia would need Perplexity to keep growing very quickly for several years, improve its economics and defend a durable position against Google, OpenAI, Anthropic and Microsoft.

But we should also value the investment from Nvidia's side rather than pretending Nvidia is a normal venture fund.

A larger Perplexity can buy more AI infrastructure. It can expose more users to Nemotron. It can provide demanding workloads for Vera and other Nvidia products. Its engineers can work with Nvidia on model and agent technology. Nvidia also keeps the upside from the equity itself.

That combination lets Nvidia tolerate a price that might look unattractive to an investor whose only possible return comes from selling the shares later.

Perplexity valuation versus annualized revenue

Q14What could make Nvidia's Perplexity bet go wrong?

The biggest risk for Nvidia is that Perplexity becomes a popular AI product without securing a durable place between users and the underlying models.

Competition is getting harder very quickly.

Google can add AI directly to the search product billions of people already use. OpenAI can combine ChatGPT, search and agents inside one enormous installed base. Anthropic is pushing Claude deeper into professional workflows. Microsoft can put agents directly inside Office and Windows.

Perplexity has responded by moving fast into Computer, enterprise software, browsers, desktop agents and multi-model workflows. That expansion gives the company more ways to win, while also putting it into more markets where much larger companies can attack it.

Economics are another concern. Perplexity's current annualized revenue growth is exceptional, but the company still has to pay for models and infrastructure. The Information says gross margin has remained around 60%, with the same historical caveat that some infrastructure costs associated with free or trial users have been treated outside cost of revenue. That accounting choice makes the headline margin less comparable with a conventional software company's gross margin.

Legal friction has also followed the company. Publishers have sued Perplexity over copyright and content use, while Amazon has sued over Perplexity's agentic shopping technology. Those disputes could raise costs or constrain parts of the product.

And then there is the platform risk. Perplexity gets much of its value from having access to other companies' models. If OpenAI, Anthropic or Google make their best systems less attractive to third-party aggregators, Perplexity's ability to offer the best model for every task becomes harder to maintain.

At a $30 billion-plus valuation, Nvidia is paying for Perplexity to turn today's extraordinary growth into a lasting position in AI.

That outcome is plausible. It is far from guaranteed.

Q15So why is Nvidia investing in Perplexity?

Nvidia is investing in Perplexity because Perplexity is becoming exactly the kind of independent AI platform that makes Nvidia more valuable: it uses huge amounts of compute, runs many competing models, builds compute-heavy agents and has no chip business of its own.

The search engine was the entry point. The investment thesis has grown much larger.

Perplexity's annualized revenue has more than tripled from below $250 million at the beginning of the year to above $750 million. Computer now spans more than 20 models and continues moving into Windows, enterprise tools, shared workspaces and email. Perplexity plans to use Nvidia's Vera architecture and currently distributes Nemotron alongside models from OpenAI, Anthropic, Google and others.

Meanwhile, Nvidia's own problem has become clearer. Amazon, Google and Microsoft are spending billions to develop Trainium, TPU and Maia chips. The biggest cloud companies increasingly control the model, application, infrastructure and silicon layers together.

Independent companies such as Perplexity keep another route open.

When Perplexity grows, it has to keep buying infrastructure from the market. When its users choose among several models, no single vertically integrated provider controls the entire workflow. When Perplexity distributes Nemotron, Nvidia gains a route into the model layer. When Computer creates more agent work, the total amount of computing rises.