Our podium's top three, read row by row

What the Podium said, and what we changed

The Podium on our funded-ideas feed groups companies that raised for the same idea, then ranks the groups by how many companies are in them. It counts companies rather than money on purpose: one large round can make any idea look like the story of the quarter, but it cannot make ten companies appear.

The rule that makes a group a lane is simple: its members have to compete with each other. A lane is a comp set. If a founder in it would never meet the others in a customer's shortlist or an investor's pipeline, it is not a lane.

Our matcher reads the one-sentence description we write for every company and looks for the words that define each lane. For the top three, those words were too wide. "Battery" matched every company that touches a battery. "Chip" plus "AI" matched software that schedules work across chips. And "AI agents" matched the companies selling security and infrastructure to the people building agents, as well as the agents themselves.

• AI agents doing the work: 18 → 10. The 8 that left now have their own lane, "The layer under AI agents".
• Chips built for AI: 17 → 8. Three software companies that spread AI work across chips from different makers got a lane of their own. So did three that move data between AI chips using light. The other three left the lane altogether: software that models chips, a cooling coating for chips, and a chip for neural implants.
• Grid & energy storage: 14 → 6. Two battery recyclers got a lane of their own. The other six left the lane: battery materials, a production-line tester, satellite power packs, a battery-swap network for electric-vehicle riders, power electronics, and data software for renewable sites.
• Then we read the lanes that moved up the same way. Construction & project bidding fell from 9 to 6. Its two site-robot companies (self-driving technology for heavy construction machinery, and robots that do building-site work) got a lane of their own, and an engineered-wood building material left it. Defence & national security stayed at 12. That lane is defined by its customer, and every member names defence or the military as one.
• The Podium now reads: Defence & national security (12), AI agents doing the work (10), then a tie on 8 between Chips built for AI and The layer under AI agents.

How we counted

• The window is the 90 days to 6 November: 478 companies with a confirmed round. Rumoured rounds are excluded.
• One company, one vote. A company reported twice counts once, at its latest reported amount.
• Every amount is what the announcement stated (reported). We do not check it against a filing.
• Lanes overlap, so there are no percentages of the market here. A company can sit in two lanes. The only shares below are shares of a single lane's own money.
• The change is versioned. The lane definitions are lexicon v2 from today, and the Podium page prints the version it was computed from. Because the Podium is recomputed on every build, tightening a lane changes past windows too, and the page now says so.

Finding one: the layer under AI agents

Split the old agent lane in two and the halves look nothing alike.

• Agents doing the work: 10 companies, $721.5m (reported), median $8.0m. One round, Butterfly Effect (the company behind the Manus agent), is $500m of it, or 69%. Without it, the other nine raised $221.5m.
• The layer under agents: 8 companies, $65.4m (reported), median $4.8m. No round above $20m (Restate). Two are pre-seed and one is seed. One is Series A and one Series B, and three did not state a stage.
• What the layer sells. Three sell controls: what an agent may touch, what it did, and which files it may open (Eagle Cloud, Kontext, Cygrid). Two sell somewhere to run agents: cloud infrastructure (Granite) and a $3,499 desktop computer (Ghost). Three sell reliability and speed: keeping a long task running when a step fails (Restate), keeping agents behaving predictably (Shamrock AI), and making them faster at reading websites (Reduck).

What we think that shows

Our read, and it is a read: nobody buys controls for software they are only trying out. A layer of security, reliability and runtime companies raising at seed suggests that some buyers have moved agents into live use, or that investors expect them to soon. Announcements cannot tell those two apart, and we will not pretend they can.

The practical version: if you are building in the layer, your comparables are these eight, not the ten above them. Their median is $4.8m. The $786.9m that the old lane displayed was mostly one round for a Chinese agent, and it was never your number.

Finding two: chips that answer, not chips that learn

Eight companies remain in the chip lane, and each one designs a chip that runs AI models.

• All 8 describe the chip as running models, and none as training them. The words are "running", "inference", "already-trained" and, in one case, answering "rather than training".
• $2.67bn (reported) across the 6 that disclosed an amount, median $190.6m. Tenstorrent's $1.5bn is 56% of it.
• Where they are: two American, three Chinese, one Korean, one Dutch, one with no country recorded. The American pair are Tenstorrent and Groq, and Groq did not state an amount. Outside Tenstorrent, the disclosed $1.17bn went to Korea, the Netherlands, China and the unrecorded one.
• The two neighbours, now lanes of their own. Running AI on any maker's chips: 3 companies, $449.6m (reported), of which Infinigence AI is $343.75m. Moving AI data with light: 3 companies, $140.9m (reported), of which Salience Labs is $125m.

What we think that shows, on chips

Our read, and it is a read: training a frontier model happens a few times, on hardware where one supplier is dominant and the cost of challenging it is enormous. Running a trained model happens every time anyone asks it a question, so it is the job where a challenger can win on cost per answer. These eight are making that bet, not the other one.

The two neighbouring lanes point the same way. Software that lets one model run on chips from different makers, and links that carry data between chips, are both ways of not being tied to one supplier. The buyers paying for them are the ones the challengers need.

The geography is the read we hold most loosely. Our collection is English-language, so it almost certainly misses Chinese chip rounds. Three Chinese companies is a floor, not a count.

Finding three: the grid lane was mostly the word "battery"

• Before: 14 companies, $307.5m (reported). After: 6 companies, $94.8m (reported).
• The largest company in the old lane was not a grid company. Nexeon ($135.4m) makes the silicon materials that go inside batteries. It does not compete with a company installing storage on the grid, and it now sits in no lane.
• The six that remain: Amber ($56m, 59% of the lane), Certain Energy, Octave.energy, Boldr, Rebaba and Furo. Median $8.3m. Two are seed, two Series A, one later stage, and one did not state a stage.
• Two of the six store nothing themselves. Amber trades the electricity in household batteries and cars, and Boldr turns home heating into capacity the grid can call on. They stay because the lane is about matching power to the moment it is needed, and we have widened its description to say so.

Run this on any lane someone quotes at you

"Fourteen companies raised for grid storage this quarter" is the kind of sentence that ends up on slide three of a deck. Three questions take a minute and change what it means:

• Do the members compete? Read each description and ask whether the companies would meet on a customer's shortlist. Half of three lanes failed this on our own page.
• Is the total one cheque? In all three lanes, one round was 56% or more of the money once the lane was cleaned.
• What is the median? The middle round in agents is $8.0m, in the layer under agents $4.8m, and in grid storage $8.3m. The median, not the total, is a benchmark for your own round.

Limits, stated plainly

• The descriptions are ours. Each company's one sentence is our summary of its announcement, and the matcher reads only that. "Built for running trained models" is how we summarised eight announcements, not eight technical specifications.
• What competes with what is a judgement. Every company we moved is named above, so you can disagree with a row rather than a percentage.
• We changed the instrument after looking at its output. The new definitions are dated and committed, so the next 90 days will test them on companies we have not seen.
• Small numbers. No lane here holds more than 12 companies. These are tallies: one extra company moves a median, and a share of a lane's money is a share of a handful of rounds.
• Stage and country are often missing. Five of the 10 agent companies and 5 of the 8 in the layer under them have no country recorded, so we make no geographic claim about either.
• One window, no trend. Ninety days is a photograph. Nothing here is rising or falling, because we have not compared it with anything.
• Amounts are as reported, from announcements, and two of the eight chip makers disclosed none.
• This continues an earlier study. "AI, robots, energy, space: one supply chain, four cheques" (28 August) found that a sector's total is mostly a few cheques. This post applies the same check one level down, to the lanes inside the sectors.