← ResearchRun · Resident Inference9 min readConcept

What should a company's AI box look like?

We run NVIDIA hardware inside businesses today and it works. But a company's day asks different things of a box than one person's desk does. We worked out what, and then had some fun with the shell: three concepts for what our dream hardware might look like.

Explore the three concepts → Turn them, light them up and change their finish; best on a phone held sideways or on a desktop. The reasoning behind them is below.

Most of what we build runs on NVIDIA hardware inside a customer's building: a desk-side DGX Spark, or two of them linked by a single cable. We think these are remarkable machines. A Grace Blackwell superchip with 128 GB of unified memory and a petaflop of FP4 compute, in 1.2 kg, that holds a serious model and runs it all day without a sound. Everything on this site that says "in your building" is running on one, and NVIDIA's larger Blackwell systems are where we go when a customer needs more.

We also work with Supermicro, and not only on paper. Earlier in 2026, Pentatonic and Supermicro built custom hardware for our retail kiosk, which is now being installed in retail stores in the US. So when we think about what a box for a company should be, we are thinking with the two partners who would build it.

So this is not a complaint about the hardware. It is a question about its shape. A workstation is built to answer one person quickly. A company is a hundred people and, increasingly, a few hundred agents reading on their behalf, all day. Those are different jobs, and we wanted to know what a box built for the second job would look like, inside and out.

The first half of this post is arithmetic. The second half is a design exercise. All of it is a concept: the numbers come from NVIDIA's specification sheet and from maths we show, not from a machine we have run, and the box is a drawing.

What one superchip gives you

NVIDIA publishes the figures for the GB10 Grace Blackwell superchip at the heart of a DGX Spark: 128 GB of unified memory, 273 GB/s of memory bandwidth, up to 1 PFLOP of FP4 compute with sparsity, 20 Arm cores, a 200 Gb/s ConnectX-7 network interface, up to 4 TB of self-encrypting NVMe, and a 240 W power supply. Up to four systems can be joined through that network interface.

Those four numbers, memory, bandwidth, compute and network, set the shape of everything a box can do. Here is how.

What a team's day asks of a box

Does the model fit, and can two stay loaded? Memory capacity decides which models a box can hold at all. One superchip holds 128 GB. Two, linked, hold 256 GB, which is enough for a 763-billion-parameter open-weight model from DeepSeek packed to 203 GB: that model runs on two superchips in our lab today. But a company does not want one model. It wants a fast one for people and a deep one for hard work, both loaded, so nothing waits for a model to start. That needs capacity for both at once.

How fast does one answer come out? Generating text is bound by memory bandwidth, not compute: every token means reading the active weights once more. The ceiling is simple to estimate. A sparse model with about 6B active parameters at 4-bit precision reads about 3 GB per token, so one 273 GB/s superchip tops out somewhere around 90 tokens a second for a single stream. The 763B model, with roughly 40 to 50B parameters active per token, reads 20 to 25 GB per token across a pair's combined 546 GB/s, so about 20 to 25 tokens a second. These are ceilings from the specification sheet; real engines land somewhat below them.

How many people at once? This is where a box built for a team earns its keep. Because generation is bandwidth bound, serving a second stream in the same step costs almost nothing extra until the compute runs out. Batching turns one person's 90 tokens a second into many people's answers at little cost each. Desk hardware rarely gets to use this. A company's box uses it all day.

How fast can it read? Agents mostly read: long documents in, short answers out. Reading (prefill) is compute bound and runs on the FP4 tensor cores, which is the one place a superchip's peak compute matters. A rough estimate: at about 150 TFLOPS effective per superchip, a pair running a 13B-active model reads on the order of ten thousand tokens a second, and a pair running the 763B model about three thousand. Over a working day that is on the order of a billion tokens read. We give the arithmetic so you can check it.

Is the person still first? The number a team actually feels is time to the first word when the box is busy. Our serving already puts people ahead of background work: first word in about 0.2 s, even with background agents running. A box built for a team keeps that true at a hundred people and a few hundred agents, which means the fast model must never be queued behind the deep one. Separate hardware for each is the simplest way to guarantee it.

What does it remember between agents? Company context (policies, shared documents, the memory) can be computed once, kept warm, and reused by every agent that starts afterwards. That trades storage and memory for compute, and it is why a company's box wants far more NVMe than a desk does.

Who looks after it? A workstation assumes someone is sitting at it. A company's box runs unattended: it needs a management controller that can power-cycle a stuck part remotely, a vault that is encrypted from the first boot, and a way for support to get in that the customer can see and stop.

What changes if you build for the team

Put those requirements together and a shape falls out. We call the concept Quartet (a working name; Pentatonic means five notes, and the box plays four).

  • Four superchips on one fabric. Four GB10s, each on its own slide-out sled, joined by a 200 Gb/s switch on a midplane designed with Supermicro. Design targets: 512 GB of unified memory, about 1.09 TB/s of aggregate bandwidth, up to 4 PFLOPS of FP4 compute, 80 Arm cores, 32 TB of NVMe. NVIDIA already supports four-system clusters through a switch; the custom part is putting the switch and the management on one board inside one designed chassis.
  • Two minds. The four run as two pairs. The deep pair holds the 763B model. The fast pair holds the fast model, the memory engine and extraction. People get the fast mind first; hard work escalates to the deep mind. Both are always loaded, and neither can queue the other.
  • Resident context. Shared company context is computed once, kept in memory and spilled to NVMe, so the 300th agent of the day starts where the first one finished.
  • Built to be left alone. A management controller (we call it the Steward) runs power, sealing and the status light. The vault is encrypted from the first power-on, and its recovery key is split across three cards, any two of which restore it. A physical switch on the back cuts the outbound path. Support sessions are time-boxed, logged and shown on the face of the box while they are open.
  • Serviceable. Sleds slide out of the back. A failing one is swapped by anyone; the box runs on three in the meantime. When the next superchip generation arrives, new sleds go into the same box.

Two sleds in the same chassis is a smaller sibling (256 GB); two Quartets linked through the rear ports make 1 TB.

How we would prove it

The method is settled, and it is the kind we publish before we run: replay a 100-person company's day, with people asking questions through the morning and afternoon peaks and a few hundred agent runs underneath, against three configurations. One Spark pair with one model, as we install today. Two pairs running the same model, to see what more hardware alone buys. And two pairs as two minds. We would measure the time to finish the day's agent work, the p95 time to first word for people under load, tokens served, and energy per million tokens, at the wall. If the two-minds arm does not beat the doubled arm, the architecture argument above is wrong, and we will say so.

Form is a blank canvas

The inside of the box is set by physics. The outside is not. Four sleds, two fans, a power supply and a controller make a package of roughly 10 to 20 litres, and that package can take almost any shape.

Pentatonic began in materials: take-back, recycling and product lifecycles for global consumer brands, and a long-held view that hardware should be built to be repaired, refurbished and returned. So we had some fun. Working with a designer, we drew three shells for the same package, each with a different idea about what a company's AI should look like in the room.

Plasma Cube. Knowledge held inside a boundary. A 280 mm cube of polished, black-anodised recycled aluminium, laser-perforated with holes too small to see over a fine LED grid. It reads as black glass until it lights, and then a graph of the company's activity glows inside it, each team its own colour. It stays metal, so it still cools, shields and contains.

Abydos. The future, carved in. A 400 mm disc with the feel of temple stone, coarse bead-blasted recycled aluminium anodised a pale sand, its only mark engraved in the top. Five rings of light circle the wall, and bars of light fly round each one at their own pace and in their own direction: different parts of the business, sharing knowledge inside one boundary.

Panel. Lit by the work. One plain, slim block, 140 mm wide, as slim as four sleds standing on edge allow, in mirror, matte copper or space grey. Nothing lights up except the mark on the front, whose colour follows how hard the company's AI is working: white at rest, through sage, to leaf green at peak.

Each concept comes with a drawing showing that it can be built: airflow, heat, noise, service access and the fire enclosure. The materials are recycled aluminium throughout, a flax-fibre composite where a natural surface is wanted, copper lattice heat exchangers grown electrochemically at room temperature, and packaging grown from mycelium that doubles as the return kit.

Explore the three concepts → They turn, they light up, and you can change their finish. It works best on a phone held sideways, or on a desktop.

Should we make this?

This is a concept, and we would like to know whether it should be more than that.

If your company runs on hardware in its own building, or would like to: would you buy this? Which of the three would you put in the room? What did we get wrong?

Tell us: hello@pentatonic.com

Pentatonic Labs · 15 November 2026 · Concept