AMD Falls to 610.20 as Perplexity Moves Agents Onto PCs
Perplexity's move to run AI agents on a local portable computer knocked AMD shares lower, with the stock at 610.20 and down 3.24% intraday against a 0.61% drop in the Nasdaq 100.
AMD shares fell 4.5% after Perplexity said it would run AI agents locally on a portable computer, shifting some inference away from cloud tokens toward owned hardware; the stock traded at 610.20, down 3.24% from a prior close of 630.63, as of 16:27:50 GMT on Sept. 28, 2026.
AMD (AMD) dropped as much as 4.5% on Monday after Perplexity said it would run AI agents directly on a portable computer, keeping inference on the device rather than sending every request to a rented cloud endpoint. By the last trade at 16:27:50 GMT the shares were quoted at 610.20, down 3.24% from a prior close of 630.63, with an intraday range of 596.07 to 629.75.
The sell-off was specific. The Nasdaq 100 tracker (QQQ) was off 0.61% at $739.98, the S&P 500 proxy (SPY) down 0.29% at $769.09 and the Dow tracker (DIA) down 0.22% at $516.33. AMD's decline ran roughly 2.6 percentage points wider than the Nasdaq 100's on the day — a gap that says the market treated this as a company-level story about where AI work physically happens, not a broad risk-off session in technology.
What changes when inference leaves the cloud
Inference is the run-time half of artificial intelligence: the moment a trained model actually answers a question, drafts an email or executes a task. Today most of that work is metered. A developer buys tokens from a cloud provider, the provider buys accelerators, and the accelerator vendor books the revenue in large, lumpy data-center orders.
Putting agents on a local machine breaks that chain in one place. As GuruFocus framed it, the portable computer shifts a portion of inference from cloud tokens to hardware ownership. The user pays once for silicon instead of continuously for compute. For a chipmaker, that is not automatically a loss — the silicon still has to be sold — but it moves the revenue from a channel with visible hyperscaler order books into a consumer replacement cycle that is slower, more seasonal and harder to forecast.
Why the market read it as a Ryzen problem
The mix question sits at the center of AMD's valuation. Its data-center accelerator franchise is the reason the stock carries an AI multiple at all; its Ryzen client business is the older, steadier, lower-margin half. A credible on-device agent product pushes demand toward the client side while implicitly capping some of the incremental cloud inference that funds the accelerator story.
The lead framing — that this widens Ryzen's software test — is the sharper point. Client silicon competes on more than raw throughput. It competes on whether developers ship models that target the on-chip AI engine, whether drivers and runtimes are stable, and whether a mainstream application like an AI agent runs well enough on a laptop that buyers notice the difference. That is a software distribution problem, and it is the one AMD has historically had the hardest time solving relative to the incumbent GPU ecosystem.
So the bear case behind Monday's move is not "fewer chips sold." It is: the chips that get sold shift toward a segment where AMD has to win on a software stack it does not fully control, and away from a segment where hyperscaler purchase commitments give investors something to model.
The bull reading nobody priced on Monday
There is a symmetric argument. If AI agents become a reason to replace a laptop, the client PC market gets a demand catalyst it has not had since the pandemic pull-forward. AMD sells into that market with a product line already positioned around integrated AI acceleration. A high-profile software partner building agents for local execution is, on that reading, free validation of the premise that consumer silicon needs an on-chip neural engine at all.
The two cases are not mutually exclusive over any reasonable horizon. Training stays in the data center regardless. Large-context, multi-step, tool-using agents will keep hitting the limits of what a battery-powered device can do, and will fall back to the cloud. What is genuinely in play is the split of routine, low-latency, privacy-sensitive inference — and that split is where the margin math gets decided.
Reading the tape against the range
The day's trading tells you something about conviction. AMD opened the session's range near its prior close at 629.75 and sold down to 596.07 before recovering to 610.20 at the last print — a full-day swing of 33.68 points between the low and the high. The recovery off the low is worth noting: buyers stepped in below 600 rather than letting the headline run. That pattern is more consistent with a re-rating debate than with a disclosure of damage.
The gap between the reported 4.5% drop and the 3.24% decline at the 16:27:50 GMT print is simply the difference between the intraday low and where the shares settled as the session ran on. Anyone using the headline number is quoting the worst moment of the day.
What to watch from here
- Attach rates, not announcements. The question is whether local agents ship to enough devices to register in unit demand, or remain a flagship demo on one portable computer.
- Developer targeting. If model developers compile and optimize for AMD's client AI engine as a first-class target rather than an afterthought, the software test starts going AMD's way.
- Data-center order commentary. Any sign that cloud customers are moderating accelerator commitments because inference is migrating to endpoints would turn Monday's thesis from speculative into measurable.
- Peer behavior. A durable shift toward the edge would show up across the client silicon complex, not in one stock. A move confined to AMD is a sentiment event.
The immediate significance of a single product decision by one AI company is easy to overstate. What is harder to dismiss is the direction: every quarter, more of the AI stack is being tested on hardware the user already owns. For a company whose premium rests on selling the metered version of that work, each such test invites the market to re-check its assumptions. Monday's 3.24% decline was that re-check, not a verdict.
All prices cited are as of the last trade at 16:27:50 GMT on Sept. 28, 2026, while the market was open.
Key facts
- AMD last trade: 610.20, -3.24% (as of 16:27:50 GMT, Sept. 28, 2026)
- Prior close / day range: 630.63; intraday 596.07–629.75
- Headline decline: Stock dropped 4.5% at its intraday low
- Nasdaq 100 (QQQ) same session: $739.98, -0.61%
Frequently asked questions
Why did AMD stock fall on September 28, 2026?
Shares dropped 4.5% at their worst level after Perplexity said it would run AI agents locally on a portable computer. That shifts some inference workload from metered cloud tokens toward hardware the user owns, raising questions about AMD's data-center accelerator demand and putting more weight on its Ryzen client software ecosystem.
Where did AMD shares actually trade?
At the last trade recorded at 16:27:50 GMT on September 28, 2026, AMD was quoted at 610.20, down 3.24% from a prior close of 630.63. The intraday range ran from 596.07 to 629.75, meaning the shares recovered from their lows before the print.
What is AI inference and why does it matter here?
Inference is the run-time stage of an AI model — when it answers a question or performs a task, as opposed to being trained. Most inference today runs in cloud data centers and is billed per token. Moving it onto a local PC replaces that recurring charge with a one-time hardware purchase.
Is on-device AI bad for AMD?
Not necessarily. It could reduce demand for cloud accelerators, which carry AMD's AI valuation premium, but it could also drive a PC replacement cycle that benefits Ryzen client chips. The unresolved issue is whether developers optimize their models for AMD's on-chip AI engine rather than competing platforms.
How did AMD perform against the broader market that day?
AMD's 3.24% decline was far steeper than the benchmarks. The Nasdaq 100 tracker fell 0.61% to $739.98, the S&P 500 proxy fell 0.29% to $769.09 and the Dow tracker fell 0.22% to $516.33. The gap indicates a company-specific reaction rather than a broad technology sell-off.
What would confirm or refute the bearish thesis?
Watch whether local agents reach enough shipped devices to move PC unit demand, whether model developers treat AMD client silicon as a primary optimization target, and whether cloud customers signal any moderation in accelerator purchase commitments. A shift confined to one stock is sentiment; a sector-wide move is structural.
