Research · 2026-08-26

Tokens per watt: reading Apple's August like a garage owner.

On August 25 Apple announced the M6 and M5 Ultra, a new Mac mini, and a new Mac Studio. We read the three announcements the way we read everything: as people who buy these machines, put a 🤫 Private Agent on them, and hand them to a person who will run that agent day and night in their own home. Here is what the numbers say, what they do not say, and which configurations we are most excited to ship as the 🤫 Puppy One desktop line.

01

What Apple shipped, in Apple's numbers.

Four desktop machines now span from $899 to the largest unified memory Apple has ever sold. The M6 is Apple's first 2nm chip, with a dual 16-core Neural Engine and Neural Accelerators in every GPU core. The M5 Ultra is Apple's first quad-die chip: up to a 36-core CPU, an 80-core GPU Apple calls its most powerful ever, up to 512 GB of unified memory, and 1.2 TB/s of memory bandwidth. Apple's own framing is the striking part: the Mac mini page talks about always-on agentic computing and creating AI agents that automate daily tasks, and macOS 27 brings what Apple calls a profoundly more capable Siri AI, with App Intents exposing native-app actions to agents. The machine Apple just described is the machine we have been building for.

02

The physics of a token, and why agents change the math.

A language model answers in two phases, and they are bounded by different silicon. Reading your prompt (prefill) is a compute problem: the chip crunches the whole context in parallel, so it is bounded by math throughput - exactly what Neural Accelerators in every GPU core speed up, and exactly where Apple's biggest published multipliers land: 4.8x on the M6, 4x on M5 Pro and M5 Ultra, 3.9x on M5 Max, generation over generation. Writing the answer (decode) is a memory problem: for every single token, the chip must read the model's active weights once, so sustained tokens per second is bounded by memory bandwidth divided by the bytes each token touches. No amount of extra compute makes decode faster once bandwidth is saturated.

Chat is decode-heavy: you write a line, the model writes a paragraph. An agent is the opposite. It re-reads its tools, its instructions, your preferences, the page it just fetched, and the result of every action, over and over, and emits short decisions. Agentic work is prefill-heavy, which means Apple just shipped its largest gains precisely where a personal agent spends most of its time. That is the single most important sentence in the three announcements, and Apple never quite says it out loud.

03

Bandwidth per dollar: the estimator, stated in full.

Our decode estimate is deliberately simple: tokens per second is about 0.7 times memory bandwidth divided by the bytes of active weights per token, where int4 quantization puts an 8B model near 4.5 GB, a 30B model near 17 GB, and a 70B model near 40 GB, and 0.7 is the fraction of theoretical bandwidth real inference sustains. It ignores KV-cache traffic, speculative decoding, and mixture-of-experts sparsity - all of which help - so it is a floor, not a promise. By that floor: the Mini decodes an 8B model around 26 tokens a second, the Mini Pro around 47, the Studio runs a 70B model around 10, and the Studio Ultra around 21, with room for far larger mixture-of-experts models to sit fully resident in 512 GB.

Divide Apple's bandwidth by Apple's price and the value picture is unambiguous: the Mac Studio M5 Max delivers about 0.25 GB/s of bandwidth per dollar - the best in the lineup, ahead of the Ultra's 0.22, the Mini's 0.19, and the Mini Pro's 0.18. For pure sustained generation per dollar, One Puppy Studio is the bargain of the four, and it is not close. What the Ultra buys instead is capacity: it is the only machine where frontier-class open-weight models - Apple's own words - live entirely in your house.

04

Tokens per watt, done honestly.

Here is what Apple did not publish: a single watt. No power figures appear in any of the three announcements. The internet will fill that silence with guesses; we will not, because tokens per watt is our headline metric and a headline metric you invented is a lie with units. What we can say from architecture is why we expect Apple to win it at the desk: unified memory means weights are read once from low-power DRAM sitting on the package, not shuttled across a PCIe bus into a discrete GPU's power-hungry memory; the efficiency cores carry the always-on idle that dominates a 24/7 agent's day; and 2nm is the most efficient process Apple has ever shipped.

05

The configurations we are most excited about.

One Puppy Mini is the one that matters most. At $899 for Apple's first 2nm chip with a dual Neural Engine, the always-on personal agent stops being a luxury purchase: it is the price of a good bicycle, drawing desk-lamp class power on a shelf, running the errands of a household around the clock. The mission says every American; $899 is what the door to sovereignty costing less than a phone looks like.

One Puppy Studio is the enthusiast's honest answer - the bandwidth-per-dollar champion with 70B-class models resident. And One Puppy Studio Ultra is the ceiling: 512 GB of unified memory means the best open-weight models in the world, whole, private, in your garage - and Apple states four of them cluster to 3x the inference of one. That is the garage grid, assembled from parts anyone can buy at an Apple Store.

One honest note on the word partner. We build on Apple's platform, in public admiration of this hardware and of the software underneath it - Foundation Models, App Intents, the new Siri AI, and Private Cloud Compute's verifiable privacy posture. We buy the same machines you can. Apple has not endorsed us, and nothing here claims otherwise; the respect flows one direction, and it is ours to give.

06

Predictions we can be wrong about.

A thesis you cannot falsify is marketing. So: we predict the M6 Mac mini sustains at least 20 tokens a second on an 8B int4 model at under 40 watts at the wall. We predict the M5 Max Studio sustains at least 8 tokens a second on a 70B int4 model. We predict prefill throughput per watt on the M6 beats any consumer discrete-GPU system at its price. Puppy Bench will print the truth next to each prediction, including the ones we miss.