Pinned Tweet
Jason Li at Epoch AI wanted to know how much agent work the world's AI chips can do. Counting their arithmetic kept giving the wrong answer.
Picture a coding agent hauling 200,000 tokens of history into every request. That's Epoch's own example. For work like that, Epoch says, memory and the links between chips can matter more than the math.
So Li built a new ruler: how many agents like that a chip can serve at once. On it, Nvidia's GB300 scores about 25 times an H100. On raw math speed it's 2.5 times. Judge it by the math and you'd undersell the newer chip tenfold.
Yesterday Li ran every AI chip sold to date through it. Raw compute grew 2.6x in a year. Agent capacity grew 4.6x.
The next step is smaller. Epoch models 8x per GPU from H200 to GB300, and 2.2x from GB300 to Vera Rubin. My rough math: about a quarter of the last jump.
Epoch says the Hopper-to-Blackwell switch already captured many of the largest gains. The next 4.6x comes mostly on pallets.
Chart: ours, data via Epoch AI (CC BY)
1
1
75
The day after NASA lost its first Echo balloon satellite to a failed rocket, a photographer at Langley was in the shop with the men packaging the next one. NASA's print is dated May 14, 1960.
Echo was a 100-foot Mylar balloon, half a thousandth of an inch thick, built to bounce telephone, radio and TV signals across continents. The May 13 launch failed when the new Delta rocket's second stage malfunctioned, according to the official report on U.S. space activities for 1960. Echo 1 made orbit on the next try, August 12.
That Thor-Delta cost about $92,000 for every kilogram it could lift to low orbit, in 2024 dollars. The figure comes from a dataset the economists Alessio Terzi of Cambridge's Bennett School and Francesco Nicoli published in PNAS Nexus this year, covering more than 4,400 launches from 1960 to 2025. Averaged across every rocket that flew, a kilogram to orbit cost $87,023 in 1960 and $3,868 in 2025.
Their fit says the cost falls about 21% each time the total mass ever launched doubles.
They are careful about one word. These are their estimates of what each rocket costs to build, with development left out, and as they write, "prices are what will be charged to those wanting to access space."
The whole spreadsheet is free to download from Terzi's website.
Photo: NASA Langley, May 14, 1960.
9
On December 16, 1954, most of the GE research lab had already left for Christmas, and Tracy Hall was alone with a pressure chamber his company had refused to pay for.
The chamber needed about $1,000. GE was already buying a $125,000 press for its diamond project, but according to the Los Angeles Times it turned Hall down and wouldn't give him official machine-shop time either. A friend in the shop built it after hours, and a former supervisor talked the company into buying the tungsten carbide. Until then Hall had been working on a 35-year-old hydraulic press that leaked so badly he wore rubber boots.
When he opened the new chamber, he later said, "My hands began to tremble; my heart beat rapidly; my knees weakened and no longer gave support." He had caught the flash of dozens of tiny crystals, the first diamonds anyone made by a repeatable process.
His bosses suspected he was exaggerating, so they repeated the run with Hall out of the building, and the diamonds came out anyway. GE announced the feat on February 14, 1955. Hall's reward was a $10 savings bond. "Big deal," he said later.
His method, high pressure and high temperature, now dominates lab-grown diamonds of one carat and under. In his review of 2025, the diamond analyst Edahn Golan reported that the growers using it stopped cutting prices once rough reached around $10 a carat.
At that floor, the savings bond GE handed Hall for the first diamond would buy about a carat of rough, in nominal dollars, before anyone cuts it.
Photos via Wikimedia Commons: Hall with his later tetrahedral press, 1960s, Brigham Young University News Bureau (public domain); GE Research Laboratory, UpstateNYer (CC BY-SA 3.0); HPHT synthetic diamond, Marie-Lan Taÿ Pamart (CC BY 4.0); synthetic diamond crystal, Ludvig14 (CC BY-SA 3.0).
1
38
GE's February 1955 news release implied the diamonds came out of the company's expensive new press, the LA Times obituary noted. Hall left for BYU and designed the tetrahedral press in that photo, which got around GE's patents entirely.
latimes.com/archives/la-xpm-…
10
In February 2004, Jeffery Schloss at the NIH put his name on a request that read like a typo: a human genome for about $1,000.
Reading one still cost $10 million to $50 million, by the request's own count, so he was asking for a price cut of four orders of magnitude. The request guessed it might take ten years.
NHGRI has tracked what its funded centers pay per genome since 2001, so we can grade the guess. At the ten-year mark, January 2014, the number was $4,008, still four times the goal. It finally slipped under $1,000 in early 2019, five years late, and kept going to $525 by 2022.
The stretch of the chart I can't stop staring at is 2008, when the centers swapped capillary machines for next-generation sequencers. The cost went from $7.1 million to $1.4 million in six months.
All in, $95.3 million to $525 is about 180,000 times cheaper. Moore's Law pace over the same 21 years would have gotten you roughly 1,300 times (my math, halving every two years).
Chart: ours, data via NHGRI
14
Price of the Future retweeted
AIで建設工程を5,000回シミュレーション、遅延要因を発見
🔹天候・作業員の生産性・資材納入遅延を変えて検証
🔹97%の試行で外部資材置き場が完工日を左右
🔹128通りの変更案を比較。2施策の併用で工期内完成確率が1%未満から46%に
実験:Timothy Fairley氏
情報元はリプ欄。
SA Water drilled boreholes before it dug. The granite under the seafloor at Billy Lights Point still came in denser than expected.
The job is a desalination plant outside Port Lincoln, South Australia: up to 5.3 billion litres of drinking water a year for the Eyre Peninsula. When the state approved it in 2024, the ABC reported a price of A$330 million. In July 2025 the state revised it to about A$470 million, after a redesign that put the seawater pipes in a tunnel under the shoreline, plus inflation.
Then a 60-tonne boring machine named Janette went under the beach. By July this year she was about 80% of the way through the 440-metre tunnel and crawling. "The density of incredibly hard layers of granite rock below the seafloor is greater than expected," SA Water's chief executive David Ryan said. Crews adjusted her cutting disc and brought in more tunnelling crew.
Janette broke through in October. The pipeline on land is built, and SA Water's Peter Seltsikas told the ABC the plant is "substantially complete." First water still slid from the end of 2026 to the end of June 2027, because everything at sea waits on that hole.
The granite's bill hasn't been written yet. SA Water says the delay means higher costs, with a revised budget once contract processes with its delivery partners are done.
A borehole tells you about the rock it touched. The rock between the holes sends its invoice later.
Photo: Jacqui Barker, CC BY 2.0, via Wikimedia Commons
1
35
The tunnel is only the sleeve. Next, SA Water says, crews install a 450-metre pipe to draw seawater in and a 1,040-metre outfall to send the brine back out.
sawater.com.au/news/ep-desal…
7
To keep growing 4.6x a year, each year's new chips must add 3.6 times all the capacity sold before them, Epoch says. Chips sold, not chips running. Capacity, not price.
epoch.ai/data-insights/agent…
6
Lizhi Sun, a PhD researcher at UNSW Sydney, went after the silver on the back of a solar cell.
The Silver Institute's World Silver Survey says silver went from 8-10% of a solar cell's cost at the start of 2025 to over 20%.
In a paper out Oct 27, Sun and colleagues swapped a tenth of the silver powder in a cell's back-contact paste for nickel. With the right nickel powder, metal cost on those lines fell 6.8%, about $561,000 per gigawatt of cells by the paper's math. Efficiency slipped 0.024 points.
Then they tried 15% nickel. The paste spread as it printed, and the lines widened from about 21 microns to 27. Each cell took 17% more paste, so it used almost as much silver as before.
At the paper's assumed prices, nickel costs about 1/140th as much as silver. It still came within 0.3% of saving nothing.
Image: Sun et al., Sol. Energy Mater. Sol. Cells (2027), CC BY 4.0, cropped
1
54
The ITRPV industry roadmap says June's silver price put a typical TOPCon cell's silver at around 2.5 cents a watt. InfoLink had the whole export cell at 4.4 cents a watt this week. Different sources and dates, so treat that as rough.
Paper (open access): doi.org/10.1016/j.solmat.202…
33
Referees never send an invoice, so peer review looks free. In 2021 a Budapest psychologist, Balazs Aczel, and two colleagues tallied the unpaid evenings anyway: over 100 million reviewer hours in 2020, which they called a floor. They titled the paper "A billion-dollar donation." Casey had a machine reread his old field on a Claude Max plan. teapawt.org/CJHandmer/status/21226…
In my ongoing quest to find the limits of Opus 5.5, I directed it to audit my old scientific field, theoretical astrophysics. We indexed about 150,000 papers, audited 50,000 for issues, and did progressively more thorough checks on thousands more. We found nearly 1000 major errors that got cited many times, most are still uncorrected. This survey enabled me to estimate the efficacy of peer review for correcting errors (not very good) and to identify dozens of open questions that anyone with spare Claude Max tokens could answer today (.md context files included).
Bottom line: AI can speed up quantitative science by >1000x with no further model improvement required. But I'm seeing no evidence of this being pushed within universities, only academics complaining about labs dropping math proofs.
Among my fellow reformed academic high agency founders, however, several are speed running entire disciplines.
This survey has implications for how science will be run in the very near future. Hint: Academic publishing is dead. It was already a very stupid way to record and promulgate knowledge, and had been so for decades, even before all that publicly funded knowledge was locked behind paywall cartels. Now AIs can generate thousands of novel, correct papers per day, it is no longer meaningful for humans to consider a research paper a "unit of science" or to even bother much with the 99.9% of science which, up to this point, was necessary drudgery. Time to discover something like General Relativity or Evolution every week. LFG!!
Enjoy: science.caseyhandmer.com/phy…
25
Picture someone at a kitchen table in April with a home solar quote, phone out, looking up what the panels themselves cost.
The search looks great. SEIA and Wood Mackenzie put the average module for non-utility projects at $0.37 a watt in Q2 2026, down 16% in a year.
Then the quote. Their modeled national average for a whole home system that quarter was $3.36 a watt, down just 1.4%.
Rough math, the panels are about 11% of that. The 16% panel drop was worth about 7 cents a watt, roughly 2% of the system, even if every cent reached the homeowner. On a hypothetical 8 kW system, that's about $3,000 of panels in a $26,900 quote.
To the person at the table, that looks padded. The other 89% is everything that doesn't make the brochure photo: the sale, the permit, the crew, the racking, the inverter, the truck, the margin. Logistics alone rose more than 15% in the same report.
The panel is the cheap part. It's just the part you can Google.
Chart: ours, data via SEIA/Wood Mackenzie SMI Q3 2026
1
30
The sale may outweigh the panels. Wood Mackenzie projects US home solar customer acquisition cost at $0.84 a watt in 2026. It's a forecast from another data set, so don't add it to the system price, but it's over twice the panel.
seia.org/research-resources/…
woodmac.com/news/opinion/us-…
12
In 1936, an airplane engineer named T.P. Wright begged his readers' pardon for writing down the number 100,000.
The Bureau of Air Commerce wanted a two-seat airplane you could buy for $700, as long as a factory could build 10,000 of them. Wright, at Curtiss-Wright, had been keeping a chart since 1922, started from two or three data points. It said every time a factory doubles the planes it has built, the average labor in each one drops about 20%.
So he priced the dream. At 10,000 planes: $2,090 in 1936 dollars, about three times the hope. He doubted anyone could get under $1,750. Then he ran his chart out to 100,000 planes and asked to be pardoned for "mentioning a figure so out of keeping with experience to date."
Nobody got a $700 plane in the garage. The chart turned out to fit a lot more than airplanes. Solar panels have ridden nearly the same slope, about 20% cheaper with every doubling of what's been installed, from $99.60 a watt in 1976 to 12 cents in 2025 (Our World in Data, panel only, 2025 dollars).
He set out to let the air out of a $700 airplane and wrote the price list for sunlight instead.
Chart: ours, data via Our World in Data and IRENA
1
47
The Bureau did crown two winning designs. One, the Hammond Y, got built, and its makers had to price it at what the Smithsonian calls "an unmarketable $3,000," in 1930s dollars. Its national demo tour drew interest and no buyers. Wright's paper: doi.org/10.2514/8.155
11
Price of the Future retweeted
Replying to @ApplePerception
Cloud People of Peru
youtube.com/watch?v=XHzEImYX…













