can I buy this chip today?

No. The team wrote the design and laid it out in 28nm, but no chip was manufactured. Every power and speed figure comes from design tools. Build version one on a small board you can already buy, and treat the paper as a signal about where power budgets are heading.

what number should I measure first?

Watt-hours per answer. Speed tells you how a box feels. Power tells you what size panel and battery you must buy, and therefore what the box costs to build and run.

why size for the wet season?

Because that is when the forecast is worst. In Ibadan the error on total sunlight was 0.12 in the dry season and 0.27 in the wet season, and the direct beam error reached 0.50. A box sized on dry season sunlight goes dark in the wet months.

how strong is the evidence behind this?

The chip paper scores 6 out of 10, held up by publication at a top conference and pulled down by having no fabricated part. The Ibadan solar paper scores 4 out of 10, and covers one site with weather records from 2005 to 2022. Neither was tested in a shop.

Sell AI answers that run on a solar panel, not the grid

The finding this rests on

A team designed a chip block for running DeepSeek models on small devices. It reports 109.4 TFLOPS per watt. The paper puts that at 19.35 times an H100 graphics card at FP8. Power draw runs from 122 to 345 milliwatts. That is phone territory. The full article is at chip design claims 19 times the AI work per watt of an H100.

Read the score before you read the claim. We give this paper 6 out of 10. It appeared at the flagship chip design conference, and many established labs are on it. But no chip was made. Every power figure comes from design software. The best energy score was taken at the slowest setting, 0.6 volts and 200 MHz, not the fast one.

So this is a bet on a direction, not on a part you can buy. The direction is the part that lasts. The limit for AI on small devices is power, not speed. As that article puts it, a model that needs 700 watts cannot live in your pocket, and a block that needs a third of a watt can.

The second thing this rests on is sunlight. Researchers in Ibadan forecast hourly sunlight from plain weather records. Their yearly error was 0.19, which means the typical miss was about 19 percent of the average. It fell to 0.12 in the dry season and rose to 0.27 in the wet season. That work is at weather data alone can forecast solar power in Ibadan. We score it 4 out of 10.

Who has this problem

Picture a farm supply shop in a market town outside Ibadan, in Nigeria. Four staff. Farmers walk in all morning with a leaf in a bag and one question. What is wrong with this, and what do I buy?

The staff know the common answers. They do not know the rare ones. The nearest person who does is a phone call away, and that call costs money and time.

Then the power goes. In many Nigerian homes an evening with no power is normal. The shop has a generator, and the generator drinks fuel. Phone data works some days and not others.

So the shop has an answers problem and a power problem at the same time. Fix one and the other still bites.

What you would sell them

Sell answers on a box that makes its own power. The box holds a panel, a battery, a small computer with a model already on it, and a local wifi point. Staff and customers reach it from any phone, with no data plan.

On Monday morning a farmer holds up a leaf. Someone photographs it on their own phone, sends it to the box over local wifi, and reads the answer. No grid. No airtime. No call.

You are not selling hardware. You are selling a working answer service, at a fixed monthly price, with the box included.

What to charge

Test it at 25 dollars a month a site, box included, on a two year term. That is a hypothesis, not a price. You are finding out whether a shop treats answers as a bill it pays every month.

Set it against something the shop already buys. One is generator fuel. You do not need our number for that, because the owner knows exactly what a week of fuel costs. The other is phone data. Ask what the shop spends on airtime in a month, then place your price beside it.

Price per site, not per question. A shop can budget a fixed monthly line. A shop cannot budget a number that moves with how busy the morning was.

How you would build the first version

Design no chip. The chip in the paper does not exist yet, and you are not going to make one. Buy a small board that already runs a model, a 12 volt battery, a panel and a charge controller.

Then measure the one number that decides everything. Watt-hours per answer. Run 100 questions through the board and log the power. That number, and not speed, tells you the size of panel and battery you must buy.

Size the solar side with free tools. The Ibadan team used ordinary weather records and PVLib, a free Python tool, to turn sunlight into a power figure for two real panels. Do the same for your town. Then size for the wet season, not the dry one. The forecast error was 0.27 in the wet season against 0.12 in the dry season, so the wet months are where you get caught out.

Do not build a fleet dashboard. Do not train your own model. Do not make your own battery pack. Build one box, put it in one shop, and watch it for a month.

The one-week test

Day one. Take one board, one battery and one panel. Run 100 questions. Write down watt-hours per answer, and how hot the board gets.
Day two. Pull free weather records for your town. Put them through PVLib. Work out the panel and battery size that keeps the box alive through a bad wet season week.
Day three. Visit ten shops in one market town. Ask one question. What happens when a customer asks something you cannot answer and the power is off? Write the answers down.
Day four. Go back to the shops that gave you a real story. Show the box working. Offer 25 dollars a month, and say out loud that you are testing that price. Ask for the first month now, in cash.
Day five. Count the months you were paid for. Three from ten is a yes. Zero is a no, and you stop.

What would kill this

The chip is a drawing. No part was made. Real silicon often falls short of what design tools promise. The comparison is also uneven. The design is drawn in a 28nm process and an H100 is built in 4nm, and a newer process alone brings a large power gain. If you sold this on the 19 times figure, you sold something that does not exist.

The cheap answer may be a worse answer. The design saves power by skipping work it judges to be near enough to work it already did. The paper does not report an accuracy score for the model after that skipping. So the size of the cost is unknown. A farm shop that gets a wrong answer about a leaf loses a crop, and then loses you.

A chip tuned to one model family may age fast. The tricks work because DeepSeek repeats itself in known ways. Models change every few months. Your box must survive that, so keep the model swappable.

The sun does not read the forecast. The Ibadan work missed by about 19 percent of the average across a year, and by more in the wet season. The direct beam error hit 0.50 in the wet season. Undersize the battery and the box goes dark in the week your customer needs it most.

The shop may not pay. Everyone likes a free answer. A monthly bill is a different question. That is the whole point of the one-week test, and you should expect a no.

Neither paper was tested in a shop. One is a design study on one workload with no chip. The other is one site in one city with weather records from 2005 to 2022. Both point the same way. Neither proves a business.

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.