analysis February 23, 2026

The Moat Test: AI’s First Great Shakeout Is Underway

A Google executive tells the world that AI wrapper startups are running out of time. Samsung answers by embedding Perplexity - an AI aggregator - directly into the Galaxy S26. And Nvidia walks into its most consequential earnings week yet. What looks like three unrelated stories is actually one.

Three stories landed over the weekend that, read separately, look like routine industry news: a Google executive warns that AI wrapper startups are in trouble, Samsung announces a Perplexity partnership for the Galaxy S26, and Nvidia walks into its most anticipated earnings week since the AI boom began. Read together, they are the same story - the first serious test of which AI businesses actually have defensible foundations and which ones were built on borrowed time.

The shakeout has started. The question is not whether it is happening but where in the stack it will hit hardest.

Google Calls Time on the Thin Layer

Darren Mowry, who runs Google’s global startup organization, told TechCrunch on Saturday that two categories of AI startup have their “check engine light” on: LLM wrappers and AI aggregators.

The warning was specific. A wrapper, in Mowry’s framing, is a company that takes an existing foundation model - Gemini, GPT-5, Claude - and builds a product layer on top of it that is “almost white-labeling that model.” The intellectual property is thin. The differentiation is mostly interface. As foundation models improve and providers build increasingly capable native features, the gap between the wrapper and the underlying product narrows until it disappears. Mowry’s assessment: “The industry doesn’t have a lot of patience for that anymore.”

AI aggregators - platforms like OpenRouter that route queries across multiple models through a unified API - face a related but different problem. Many have grown quickly because they offered something the models themselves did not: choice and convenience. But Mowry argues that users now want “some intellectual property built in” to ensure they’re routed to the right model for their specific need at the right time. A platform that routes traffic but doesn’t make intelligent decisions about where to send it is not, in his view, providing durable value.

The timing of the warning is deliberate. Google runs one of the world’s largest startup programs. It has visibility into which cohorts of AI companies are growing and which are stalling. When an executive in that position goes on record with a category-level alarm, it is not speculation. It is a reading of observed data.

What Mowry is describing is the beginning of what every technology platform cycle eventually produces: a consolidation around businesses that have built something the underlying models cannot simply absorb. The companies with vertical depth, proprietary data pipelines, customer workflows embedded in their product, or defensible distribution - those survive. The ones that bet on being the interface layer between a user and a model that the model provider is rapidly improving - those don’t.

Samsung’s Answer Is More Complicated Than It Looks

The same weekend Mowry warned about AI aggregators, Samsung announced that Perplexity - the AI search startup that is, by Mowry’s definition, partly an AI aggregator - will be the second AI agent in its Galaxy AI ecosystem, shipping with the Galaxy S26 series.

On first read this looks like a contradiction. Google says aggregators are in trouble; Samsung just signed a major distribution deal with one. But the tension resolves on closer inspection, and what it reveals is more interesting than either story alone.

Perplexity’s integration in Galaxy AI is not a simple chatbot embedding. According to Samsung’s announcement, Perplexity’s agent will work natively across Samsung Notes, Clock, Gallery, Reminder, and Calendar - accessing on-device context, acting across apps, and operating in the background. This is not an aggregator routing queries to multiple models. It is a company that started as an AI search product and has spent the last year building something that looks increasingly like a general-purpose agent runtime.

That distinction is exactly what Mowry was drawing. The aggregators in trouble are the ones that have not extended beyond query routing. Perplexity, by contrast, has built a distribution relationship with the world’s largest Android manufacturer and earned deep integration rights across the system UI. That is defensible in a way that an API layer is not.

Samsung’s motive is equally legible. The company has positioned Galaxy AI as a multi-agent ecosystem rather than a single-model bet. The Galaxy S26 will ship with at least two AI agents - Bixby, which Samsung owns, and Perplexity, which Samsung does not. The design choice reflects a recognition that no single model will dominate all use cases and that smartphone users increasingly expect to choose which AI handles which task. Samsung is becoming a distribution platform for AI agents, which means that what matters for a startup is not just being good - it is getting the device-layer deal.

That asymmetry - where hardware distribution is the new leverage point - is a direct challenge to Google’s own AI ambitions. Google Assistant, Gemini, and Search are all competing in the same device-layer space that Samsung just opened to a third party. The irony of a Google executive warning about AI aggregators on the same week Samsung hands one of them a flagship device slot is hard to miss.

Nvidia’s Earnings Are a Test of the Entire Thesis

The week’s most consequential event does not arrive until Wednesday. Nvidia reports Q4 FY2026 results after market close on February 25, and the stakes are unusually high even by the company’s recent standards.

Nvidia enters the report having delivered twelve consecutive quarters of earnings beats. Its CFO, Colette Kress, stated last quarter that the company sees a $500 billion opportunity between then and the end of 2026 across its Blackwell and Rubin GPU lines. Demand for Blackwell systems is, by multiple industry accounts, sold out through mid-2026. The setup is as bullish as it gets.

And yet Nvidia stock, along with the rest of the “Magnificent Seven,” has stalled in 2026. After a 35% gain over the past year, shares have gone sideways. The market has repriced the near-term expectation upward and is now asking a harder question: is the infrastructure spending cycle durable, or is it approaching a peak?

That question matters beyond Nvidia’s stock price. The entire AI investment thesis - the one that has justified trillions in market capitalization, billions in startup funding, and the strategic pivots of every major technology company - rests on the assumption that the demand for AI compute is structural rather than cyclical. If Nvidia’s forward guidance weakens, or if there is any softening in the language around Blackwell deployment timelines, it will be read as the first credible evidence that the cycle is closer to its top than its middle.

What to watch on February 25 is not the headline revenue number, which is widely expected to be strong. Watch the tone and specificity of forward guidance. Watch whether management discusses demand expanding into new verticals - manufacturing, healthcare, autonomous systems - or whether commentary stays concentrated in hyperscaler cloud. A beat on revenue with narrowing forward visibility is a different signal than a beat on revenue with accelerating vertical expansion. The two outcomes tell very different stories about where AI infrastructure spending goes from here.

The Common Thread

These three stories are all versions of the same underlying question: where does durable value actually live in the AI stack?

See also: NVIDIA Rubin and N1X: Rewriting the Rules of AI Hardware.

For related context, see Anthropic.

On first read this looks like a contradiction. Google says aggregators are in trouble; Samsung just signed a major distribution deal with one. But the tension resolves on closer inspection, and what it reveals is more interesting than either story alone.

Perplexity’s integration in Galaxy AI is not a simple chatbot embedding. According to Samsung’s announcement, Perplexity’s agent will work natively across Samsung Notes, Clock, Gallery, Reminder, and Calendar - accessing on-device context, acting across apps, and operating in the background. This is not an aggregator routing queries to multiple models. It is a company that started as an AI search product and has spent the last year building something that looks increasingly like a general-purpose agent runtime.

That distinction is exactly what Mowry was drawing. The aggregators in trouble are the ones that have not extended beyond query routing. Perplexity, by contrast, has built a distribution relationship with the world’s largest Android manufacturer and earned deep integration rights across the system UI. That is defensible in a way that an API layer is not.

Samsung’s motive is equally legible. The company has positioned Galaxy AI as a multi-agent ecosystem rather than a single-model bet. The Galaxy S26 will ship with at least two AI agents - Bixby, which Samsung owns, and Perplexity, which Samsung does not. The design choice reflects a recognition that no single model will dominate all use cases and that smartphone users increasingly expect to choose which AI handles which task. Samsung is becoming a distribution platform for AI agents, which means that what matters for a startup is not just being good - it is getting the device-layer deal.

That asymmetry - where hardware distribution is the new leverage point - is a direct challenge to Google’s own AI ambitions. Google Assistant, Gemini, and Search are all competing in the same device-layer space that Samsung just opened to a third party. The irony of a Google executive warning about AI aggregators on the same week Samsung hands one of them a flagship device slot is hard to miss.

Nvidia’s Earnings Are a Test of the Entire Thesis

The week’s most consequential event does not arrive until Wednesday. Nvidia reports Q4 FY2026 results after market close on February 25, and the stakes are unusually high even by the company’s recent standards.

Nvidia enters the report having delivered twelve consecutive quarters of earnings beats. Its CFO, Colette Kress, stated last quarter that the company sees a $500 billion opportunity between then and the end of 2026 across its Blackwell and Rubin GPU lines. Demand for Blackwell systems is, by multiple industry accounts, sold out through mid-2026. The setup is as bullish as it gets.

And yet Nvidia stock, along with the rest of the “Magnificent Seven,” has stalled in 2026. After a 35% gain over the past year, shares have gone sideways. The market has repriced the near-term expectation upward and is now asking a harder question: is the infrastructure spending cycle durable, or is it approaching a peak?

That question matters beyond Nvidia’s stock price. The entire AI investment thesis - the one that has justified trillions in market capitalization, billions in startup funding, and the strategic pivots of every major technology company - rests on the assumption that the demand for AI compute is structural rather than cyclical. If Nvidia’s forward guidance weakens, or if there is any softening in the language around Blackwell deployment timelines, it will be read as the first credible evidence that the cycle is closer to its top than its middle.

What to watch on February 25 is not the headline revenue number, which is widely expected to be strong. Watch the tone and specificity of forward guidance. Watch whether management discusses demand expanding into new verticals - manufacturing, healthcare, autonomous systems - or whether commentary stays concentrated in hyperscaler cloud. A beat on revenue with narrowing forward visibility is a different signal than a beat on revenue with accelerating vertical expansion. The two outcomes tell very different stories about where AI infrastructure spending goes from here.

The Common Thread

These three stories are all versions of the same underlying question: where does durable value actually live in the AI stack?

Google’s warning identifies the top of the stack - thin wrapper and aggregator layers - as the most vulnerable. Samsung’s Perplexity deal identifies device distribution as a new durability mechanism that can rescue companies that would otherwise be exposed. Nvidia’s earnings will tell us whether the bottom of the stack - infrastructure hardware - is running toward a sustained mid-cycle or a near-term plateau.

The companies that will still be standing in eighteen months are the ones that have locked in defensible positions at one of those layers: deep vertical workflows that cannot be commoditized by model improvement, device-level distribution that no API can replicate, or infrastructure dominance that only a rival chip at scale can displace. Everything between those anchor points is at risk.

The shakeout is not a collapse. It is a sorting. The AI industry is large enough and growing fast enough that even the losers in this round will find somewhere to land. But the era of building a company on top of someone else’s model and hoping the model provider doesn’t notice - that era is over. The week ahead will give us our clearest look yet at what the next era is built on instead.