AI, robots, energy, space: one supply chain, four cheques

What the chart actually says

The i-Index on our funded-ideas feed ranks sectors two ways at once: how many rounds each took, and how much money. Most months the two agree roughly. Over the 30 days to 28 August they came apart, and four sectors are the reason.

Here is the window, counted plainly. The corpus is committed to our repository and every figure below is generated from it by a script that prints the underlying rows, so this is reproducible rather than asserted. Every amount is the figure the announcement stated — we do not verify it against a filing, and we say (reported) every time for that reason.

• 164 companies raised a confirmed round. 153 of them disclosed an amount, totalling $11.99bn (reported).
• AI — 16 companies, $3.86bn (reported). 10% of the companies, 32% of the money.
• Robotics — 11 companies, $1.41bn (reported). 7% of the companies, 12% of the money.
• Energy — 10 companies, $1.65bn (reported). 6% of the companies, 14% of the money.
• Space — 4 companies, $1.33bn (reported). 2% of the companies, 11% of the money.
• Together: 41 companies — 25% of the month — holding 69% of the capital.
• Other B2B Software took the most rounds of anything: 25 companies. It holds 5% of the money, at a median of $6.43m (reported).

How we counted

Four decisions shape every number here, and one of them makes this page disagree slightly with our own chart.

• We count companies, not rounds. Our corpus keys on the company and the announcement date, so when a second outlet re-reports a round a day later it leaves a second row. In this window that happened four times — Rillet, Edgify, Cambridge Aerospace and Intropy. Counting rows gives 168 rounds and $12.41bn; collapsing them to companies gives 164 companies and $11.99bn. Rows overstate the capital by 3.5%, because a re-reported round has its amount added twice.
• So the chart and this page differ by about three per cent, deliberately. The live i-Index counts rows, because it is a verified port of the database query behind it and must stay one. We would rather show you the gap than quietly close it in one place and not the other.
• Counts are over all 164 companies; capital shares are over the 153 that disclosed. These are different populations, so the ratio between a company share and a capital share is not an average cheque and we never present it as one. Where cheque size matters we use the median.
• One company sits in exactly one sector. We use Dealroom's published eighteen-sector taxonomy, frozen and versioned — seventeen of the eighteen took a round in this window. A company that could plausibly be filed in two is filed in the one that fits best, and that judgement is ours.

The largest round in each of the four is a physical asset

The four sectors invite a story about frontier technology. The four rounds that actually built their bars are about premises, power, orbit and hardware. Not one of them is software:

• Firmus — $2.85bn (reported), Australia, in the AI bar. It builds and operates large AI data centres powered by Nvidia chips. The single biggest AI round of the month is a construction and power business.
• ICEYE — $1.15bn (reported), Finland, in the Space bar. Small radar satellites for Earth observation.
• Valar Atomics — $1.0bn (reported), United States, in the Energy bar. Compact nuclear fission reactors; its Ward 250 unit reached criticality in Utah under a US Department of Energy pilot programme.
• XPeng Robotics — $900m (reported), China, in the Robotics bar. Humanoid robots, spun out of the electric-vehicle maker.

The supply chain does not respect the sector labels

The clearest evidence that these bars are related is not in the bars themselves. It is one sector over, in Semiconductors — a bar the chart does not even highlight.

Eleven companies raised there this month, and seven of them name AI in their own one-line description: chips to replace general-purpose GPUs for inference, chips that connect servers inside AI data centres, AI processors for edge devices, photonic chips for AI data centres, and a diamond coating that stops AI chips overheating. Those seven hold $656.5m of the bar's $699.1m — 94% of it (reported).

That figure does not rest on one round, which is what makes it the sturdiest number on this page: the two largest are OLIX at $312m and Xsight Labs at $300m (reported), independent of each other.

The same crossing shows up inside the AI bar itself. Of the sixteen AI companies, we read eight as selling infrastructure rather than a product — compute capacity, workload routing, servers, training data. One of them, AI Infrastructure Capital AG ($18.2m reported), describes its business as buying and operating AI servers powered by renewable energy and renting out the capacity. That is a single company sitting in the AI bar, the Energy bar and the Semiconductors bar at once. It can only be filed in one.

What we think that shows

Our read, and it is a read: these are not four sectors having a good month. They are four stages of one build-out, and the sector taxonomy is cutting across it rather than describing it.

The order runs roughly like this. Compute needs power, and the power being bought here is being bought as an industrial input, not as a climate position — a fission reactor, thermal batteries supplying steady power to factories and grids, servers explicitly sited on renewables. Robotics is where model capability turns into a physical product, which is why AI money is visible on its supply side too: one AI-bar company in this window sells data infrastructure for training embodied-AI robots, another a spatial-mapping network described as powering physical AI and robotics. And space is the same instinct at maximum capital intensity — it needs the silicon, the power and the autonomy, and it cannot buy any of them off a shelf.

If that reading is right, the useful consequence for anyone allocating is that these four bars should not be treated as four independent bets. They share suppliers, they share input costs, and a constraint in one — grid connections, advanced packaging capacity, a single chip vendor — reaches all four.

What would tell us we are wrong: the four sectors moving independently of each other over several months, or the Semiconductors bar losing its AI concentration while the AI bar keeps its size. We have one month. We cannot see either yet.

Where the story breaks

Everything above describes composition. It says nothing about how broad any of it is — and this is the part a striking chart hides.

The four largest rounds in the entire window total $5.90bn (reported). That is 49% of all the capital we recorded, held by 2.4% of the companies. They are Firmus, ICEYE, Valar Atomics and XPeng Robotics — and all four sit inside the four bars, one apiece.

So the 69% figure is not describing a market. It is describing four decisions, taken by four investment committees, in one month. Remove those four rounds and the same four sectors are 23% of the companies and 38% of the capital — still over-indexed, at a little over one and a half times their share of companies, but nothing like the picture the chart paints.

The concentration inside each bar says the same thing from closer up. One round is 74% of the AI bar, 64% of Robotics, 61% of Energy and 87% of Space. Take the top three in each and you have 91%, 89%, 97% and 99% of those bars. There is very little behind the leaders.

The clearest example of why a total misleads

Our own hand-read of the AI bar is a good demonstration, so we will use it against ourselves.

Split the sixteen AI companies into infrastructure and applications and 78% of the bar's money is infrastructure — which reads as a strong claim about where AI capital is going. Take out the single largest round and it inverts: infrastructure is 16% and applications are 84%.

Both numbers are arithmetically correct. Only one of them is a fact about sixteen companies; the other is a fact about one. The honest version of the sentence is the narrow one: the biggest AI round of the month was a data centre, and that one round is most of what the AI bar is measuring.

At seed, the premium almost disappears

This is the number that matters most if you are the one raising, and it is the least visible on any chart.

Across all stages, the median round in these four sectors is $17.7m against $8.5m everywhere else (reported) — roughly 2×. That gap is what everyone quotes, and it is real at the top of the market.

Narrow it to pre-seed, seed and Series A and it very nearly closes:

• AI, Robotics, Energy and Space, early stage: median $7.50m (reported), across 18 rounds.
• Everything else, early stage: median $5.96m (reported), across 54 rounds.
• A premium of 26%, not 109%.

What that means at the table

Our read: capital intensity in these sectors is a late-stage phenomenon in this data, not an early-stage one. A first cheque into a robotics company is priced much like a first cheque into anything else, because at that point the company is a team and a plan in both cases. The divergence arrives later, when the hardware has to be built, the power contracted and the fab time bought.

The practical version, for a founder: the headline is not your comparable. A $900m humanoid-robotics round and a $2.85bn data-centre round are not evidence that the market will pay a premium for your seed. Eighteen early-stage rounds in these sectors say the number is around $7.5m, and the sample is small enough that you should treat even that as a range.

This is the same discipline we apply to every valuation we build: a median describes a population, a total describes whoever wrote the largest cheque in it, and only one of the two is a benchmark.

One more divergence, one level down

The deals-against-dollars split repeats itself geographically, and it is not flattering reading for a British audience.

Of the 41 companies in these four sectors, 11 are British — more than any other country. They hold $187m of the $8.25bn (reported): 2.3%. The capital sits in Australia, China, Finland and the United States, in four rounds.

Our read: Britain is originating this kind of company at a rate that stands comparison with anywhere, and financing it at a different order of magnitude. One month cannot establish that, and we are not going to pretend otherwise — but it is the shape everyone in UK deeptech describes, and it is what our first thirty days look like.

Limits, stated plainly

• One 30-day window, and no trend. Our corpus begins on 17 July, so it spans 42 days — a trailing 30-day average would need 60. So nothing in this window can honestly be called rising, cooling or rotating — there is nothing to compare it against. This is a photograph, not a trend line.
• Amounts are as reported. They come from the announcement, not from a filing. Where a company disclosed nothing we count it in the company totals and leave it out of every capital figure — 11 of the 164 are in that position.
• The infrastructure/application split inside the AI bar, and the AI count inside Semiconductors, are our hand-reads of each company's one-line description. They are judgements, they are ours, and every company we put on each side of those lines is named in the post or printed by the task that generates these figures — so you can disagree with a specific row rather than with a percentage.
• One company, one sector. Several companies here genuinely belong in two or three. The taxonomy has one slot, which is a large part of the point this post is making.
• Our collection is English-language and feed-driven. It over-represents the UK and Europe on count and almost certainly under-represents China, Japan and Korea — so the geography section describes what we saw, not what happened.
• Four sectors chosen after looking at the chart. We picked AI, Robotics, Energy and Space because they are what the chart made interesting, which is a selection made with the data already in view. Treat the grouping as a hypothesis worth another month, not a finding.