'The Agentic Age' Needs a Battery: Reading Summit Keynote Claims Against Phone Hardware Reality

01 What the keynote actually promised

The Day 1 keynote at Snapdragon Summit framed 2026 as the start of an agentic era: assistants that do not wait to be asked, that observe context on the device, anticipate intent, and act across apps without a round trip to the data center. The language was directional rather than quantitative. Slides showed agents handling travel rebooking, message triage, and photo curation, with the strong implication that the phone itself, not a cloud tenant, is where the work happens.

Strip the staging and three load-bearing claims remain. First, that the NPU can sustain meaningful inference workloads for hours, not seconds. Second, that sensing pipelines, wake words, context monitoring, ambient listening, can idle at almost nothing on the always-on domain. Third, that all of this fits inside a battery budget a user would call all-day. Only the first claim is a silicon claim. The second and third are claims about duty cycles and chemistry, and those do not keynote well because they come with asterisks.

Keynote claims are structured to be unfalsifiable at launch for a simple reason: no duty cycle is specified. An agent that listens continuously but reasons once per hour is a different product from one that reasons continuously, and the gap between them is not a few percent, it is one or two orders of magnitude in average power. Until a review specifies what the agent is doing minute by minute, the slide and the spec sheet are describing different machines.

02 The power math behind an always-on agent

Start from the battery, because it is the fixed constraint. A current flagship pack holds roughly 5,000 mAh at about 3.85 V nominal, or approximately 19 Wh. A user who considers 5% overnight drain acceptable is already spending about 1 Wh on background work. That is the entire envelope an always-on agent has to negotiate for before the screen turns on the next morning, and it is an estimate: pack sizes vary from roughly 4,500 to beyond 6,000 mAh across 2026 flagships.

Now price the agent. A well-built wake-word pipeline on a low-power DSP idles in the tens of milliwatts, an estimated 10 to 40 mW, which is affordable. A context-monitoring agent that samples sensors and embeds text periodically lands, generously, around 100 mW average, which is about 2.4 Wh per day, roughly 13% of the pack. Push to continuous multimodal reasoning at even a 10% duty cycle over a 3 to 5 W NPU burst and the average approaches 300 to 500 mW, which is 7 to 12 Wh per day. That is not an all-day feature; that is half the battery before the first video.

The honest formulation is therefore not that phones cannot run agents, but that agents get a power allowance of roughly 10 to 20% of the pack before users revolt, and everything else is engineering to fit inside it. The chart below arranges the estimated states on one scale. The distance between the wake-word bar and the continuous-agent bar is the entire business argument for the split architecture described later in this article.

03 Thermal ceilings: why burst benchmarks mislead

A phone has no fan. Every watt the SoC consumes becomes heat that must pass through the chassis to still air, and skin-temperature limits, about 40 to 43 degrees C where the device touches a hand, set a sustained ceiling typically estimated near 3 to 5 W for a large flagship and closer to 2 to 3 W for a compact one. Within that ceiling, short bursts are cheap: the chassis soaks the heat. Launch-day demos and benchmark runs live entirely inside the soak window.

The soak window is minutes, not hours. Sustained-load tests routinely show 20 to 40% performance falloff after 15 to 30 minutes as governors trade clock speed for temperature, and these numbers vary by chassis design, so treat the range as estimated. For an agent, throttling does not look like a lower benchmark score; it looks like deferred reasoning, queued requests, and a silent hand-off of the hard parts to the cloud. The feature does not fail loudly. It quietly becomes the hybrid product the keynote did not describe.

Vapor chambers and graphite sheets, now standard in gaming-leaning flagships, spread heat but do not remove it. Spreading reduces hotspots and raises the burst window somewhat, but the steady-state removal rate is set by chassis surface area and the environment. This is why two phones on identical silicon post different sustained scores, and why an agent workload, which is by definition sustained, is a worse-case load for the thinner of the two devices.

04 Battery chemistry has not kept pace

The quiet fact under every agentic roadmap is that the energy tank barely moves. Lithium-ion specific energy at the cell level has climbed from roughly 200 to about 260-300 Wh/kg across a decade of handsets, a few percent per year, and these figures are estimates with real spread between cell vendors. The visible jumps in phone capacity have mostly come from spending more of the internal volume on the battery and from silicon-carbon anodes, which add an estimated 5 to 10% capacity in the same volume. Real, but incremental.

Compute-per-watt tells the opposite story: process nodes and architecture changes have delivered multiples of improvement per generation in the NPU's short life. When one curve compounds and the other crawls, the gap becomes destiny. A workload that was impossibly expensive five years ago becomes affordable not because the tank grew but because the engine got efficient enough to fit the tank. The chart below indexes the two trends to make the divergence visible.

Fast charging is the industry's compensating trick, and it is a genuine one: 60 to 100 W wired charging turns a small tank into a rarely-inconvenient tank for desk-bound users. But it trades cycle life and added heat, it does nothing on a hiking day, and it does nothing for the wireless-charging majority at night, where pads dissipate an estimated 30 to 50% of the delivered energy as heat. An always-on agent is a load that follows you outdoors; fast charging is a solution that stays home.

05 The split-brain compromise: local wake words, cloud reasoning

What actually ships, on every phone now called agentic, is a split architecture. A wake-word and light-intent model lives on the low-power always-on domain, drawing its estimated tens of milliwatts, and handles the constant part: noticing, classifying, deciding whether to escalate. Anything heavy, multi-step planning, document reasoning, image understanding at scale, is either deferred, throttled to a duty cycle, or sent to a data center. The keynote slide merges these two systems into one assistant; the battery experiences them as different products.

The split has consequences the marketing does not headline. Privacy guarantees apply only to the part that stays on the device, so the boundary between local and cloud is exactly the boundary of the privacy claim. Latency guarantees apply only to the local part, so the snappy demo is the wake word, not the reasoning. And the boundary moves with each silicon generation, which means a phone bought in 2026 will draw its line in a different place than the keynote implied.

None of this makes the hybrid design a failure. It is arguably the only design that closes, given the power math above: milliwatt sensing on device, watt-scale reasoning where power is cheap. The critique is narrower and sharper: keynote language that says on-device without saying which parts, at what duty cycle, blurs the one distinction users actually care about, and blurring it is what makes all-day claims technically unfalsifiable rather than true.

06 What buyers should watch in reviews instead of keynotes

The first number to demand from a review is an agent-duty-cycle battery test: screen off, assistant enabled, a scripted hour of realistic requests, compared against the same hour with the assistant disabled. Standard video-playback tests will not surface the agent at all, because playback is a display-and-decode workload the NPU barely touches. The delta between the two hours, likely to land somewhere between a few percent and a doubling of drain depending on implementation, is the actual price of the feature.

The second is a sustained NPU test with a curve, not a single score: throughput at 5 minutes versus 30 minutes, alongside measured skin temperature. A phone that keeps 80% of its throughput at 30 minutes is a different agentic machine from one that keeps 55%, even if the launch benchmark scores are identical. Both numbers above are illustrative shapes; what matters is that the review publishes the falloff instead of the peak.

A printable checklist, in decreasing order of information value: idle drain per hour with the agent on versus off; sustained-throughput falloff over 30 minutes with skin temperature; what still works in airplane mode, which reveals how much of the agent is genuinely local; recovery behavior after a heavy burst; and standby drain overnight with location and listening enabled. Five measurements, one afternoon, and the gap between the keynote and the hardware becomes a table instead of an argument.