How to Evaluate a Compute Company

Process node and manufacturing

The process node is the manufacturing technology used to make the chip. Names such as 3 nm or 5 nm are marketing labels and no longer measure a physical dimension. They still indicate a generation, with smaller numbers generally meaning denser and more efficient chips.

Ask which foundry will make the chip, since few can build the most advanced nodes. TSMC makes most leading AI chips. Ask also about advanced packaging and memory supply, because limited capacity there has constrained shipments.

A newer node is not automatically better. It costs more, and a design on an older node can be the right choice for a lower-cost product.

Software stack

Software often decides whether a chip succeeds. Customers need compilers, libraries, and support for frameworks such as PyTorch. They need debugging tools and documentation.

Ask how much work it takes to run an existing model. A chip that needs a team to rewrite code is a harder sell than one that runs a model with a few changes. NVIDIA's CUDA ecosystem is the benchmark every challenger is measured against.

Look for public documentation, open-source code and independent developers using the tools. Their absence is informative.

Memory and interconnect

Peak compute is the easiest figure to quote and often the least useful. Look instead at memory capacity, memory bandwidth, and the speed of links between chips.

For language models, memory bandwidth and capacity frequently set real-world speed. Our guide to the memory wall explains why. For large systems, the network between chips matters just as much.

Customers and revenue

Separate real deployment from interest. A named customer running a product in production is strong evidence. Phrases such as engaged with, evaluating, or letter of intent are weak.

Check customer concentration. A company that depends on one buyer carries more risk than one with several. Check whether the revenue comes from chips, from services, or from licensing designs, since these are different businesses.

For public companies, read the filings. Filings carry legal liability for false statements, and press releases carry much less.

Funding and finances

Look at how much money was raised, when, and from whom. Funding from experienced chip investors or strategic partners carries information. A large headline round shows investor appetite, not a working product.

Chips are capital intensive. A single advanced-node tape-out can cost tens of millions of dollars or more, and a company needs enough runway for several iterations. Ask how long the money lasts at the current spending rate.

Treat valuations as opinions about the future, not as measures of technical progress.

Tape-out status

Tape-out is the point at which a finished chip design is sent to a factory to be made. After that, the first chips come back after a wait of typically a few months. Engineers then test them in a stage called bring-up. Later come validation and then volume production.

These stages are very different. A simulation, an FPGA prototype and a test chip are not a product. Ask whether silicon exists, whether it works at the claimed speed, and whether it is in volume production. Delays are common, and a second tape-out to fix problems is not unusual.

What to be sceptical about

Be careful with the following claims.

Benchmarks against an old chip, or at a different numerical precision, are not fair comparisons. Results quoted as up to a given multiple are best cases. Performance per watt is meaningless without saying what workload and what power boundary, chip or full system, were used.

Peak figures that assume the chip is fully used rarely hold in practice. Results for an unusually sparse model, or an unusually small one, may not transfer to your workload.

Roadmaps that keep moving, and announcements of partnerships that never turn into revenue, are common warning signs. So is a lack of independent results. MLPerf entries, third-party tests and customer references all help.

A short checklist

Does working silicon exist? Who makes it? What is the delivered speed on a real model? Can an engineer run PyTorch code on it? Who is paying for it today? How long does the funding last? Who are the founders, and have they shipped chips before?

The company pages on Compute Compass list funding, founding year, location and technology tags, which is a starting point for these questions. Always check primary sources before relying on a figure.