Invest1 publisher3 min readPublished
a16z earmarks 39 cents of every dollar it raised this week for the physical AI stack
The $1.1bn Machine Age Fund is real money and still only about a tenth of one percent of what Goldman expects the world to spend on AI next year. That gap shows what venture equity can and cannot buy.
The Investor · Invest desk
What happened
- On August 31, a16z added $1.75bn to its fifth growth fund, taking it from the $6.75bn it debuted at in January to $8.5bn, an extension of the existing vehicle rather than a new fund, per TechCrunch.
- Days earlier the firm closed a $1.1bn Machine Age Fund aimed at chips, data centres, networking, storage and robotics, which is the physical layer under the models.
- The OECD counted $109.3bn of venture capital into AI infrastructure and hosting companies in 2025, against $47.4bn in 2024, or more than 42% of all AI venture money that year.
- Goldman Sachs Research forecasts about $1tn of worldwide AI-related investment in 2026, with $581bn of that in the United States.
- Both vehicles sit inside the more than $15bn of new funds a16z unveiled in January, which makes the August top-up an addition to a programme already announced rather than a new one.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- decision Widening January's vehicle instead of raising a sixth fund commits $9.6bn across two pools, roughly 10.7% of the about $90bn a16z manages, to strategies it has already described in public.
- constraint Money aimed at the physical layer cannot be deployed faster than fabs and substations allow, so the binding limit on the $1.1bn is capacity allocation, not deal flow: three years of 20-30% supply growth compounds to only 1.7 to 2.2 times.
- exposure With deals over $1bn accounting for almost half of AI venture value in 2025 by the OECD's count, a fund of this size gets marked by a handful of positions rather than by a portfolio.
- contradiction Stanford and the OECD measure different things. Infrastructure's share of the money is either 42.3% or 31.7% depending on which denominator you accept, and that changes how crowded the physical layer looks.
The extension does its arithmetic quietly: $1.75bn on top of $6.75bn is 25.9% more capital deployed under a mandate fixed in January, not a new one negotiated in August [1][2][7]. The $1.1bn Machine Age Fund is the opposite kind of act, a fresh mandate at roughly an eighth of the growth vehicle's size [3][24], and it has to underwrite things that get poured and energised rather than shipped.
a16z's stated case for the second is a supply gap it describes in hardware terms: rack compute density up 28 times from the first H100 systems to Rubin, rack power from 5 to 10 kilowatts to 100 to 250 kilowatts now, and a megawatt expected within three years [8]. Set that against the 20% to 30% annual growth the firm says hardware supply has historically managed, and three years of compounding gets you 1.7 to 2.2 times against a per-rack power requirement rising four to ten times [9]. It is not a like-for-like comparison (power per rack is not units shipped), but it is the pitch, and the pitch is that the scarce thing is now physical.
What that appetite does to prices shows up in the smaller deals. Fractile, a British chip startup, entered talks over a $6.5bn pre-money valuation after a preliminary agreement to supply about $250mn of chips, which is 26 times the contract that got it there, and a supply agreement is not revenue [21][22]. Forge Global's timing data says something similar from the other end: Anthropic, OpenAI and xAI reached $100bn valuations in roughly five years, where SpaceX, Stripe and Waymo averaged about sixteen [20]. Marks now arrive faster than they used to. That is awkward for any duration argument: hardware gets built on construction schedules, while valuations get set on funding schedules.
One reading of the split is that the growth extension is the conservative bet and the Machine Age Fund is the priced one. Gartner already has AI data centres going from 36.5% of chip revenue in 2026 to more than 53% by 2030, with memory at about $837bn of a $1.6tn semiconductor market, or 52.3% of the entire thing [14][15]. When the consensus forecast is your entry assumption, the return has to come from getting allocation rather than from being early. Goldman Sachs Research makes the same point at scale: about $1tn in 2026 against roughly $1.8tn cumulative since 2022, which leaves about $0.8tn for the four years before it, so one year outweighs the previous four combined [12][13].
What would settle it is countable. Stanford's index puts global private AI investment at $344.7bn in 2025, up 127.5%, of which $285.9bn, or 82.9%, was American, more than 23 times China's $12.4bn [16][26]. Another year near 127.5% and the hardware entry prices were cheap at 26 times a supply agreement; a year at half that, and the vehicle with the fixed thesis and the longer build cycle is the one holding the asset it cannot re-underwrite.
What to watch
- Whether a16z follows the extension with a sixth growth fund rather than another top-up, and at what disclosed size.
- Whether Fractile's talks close at the number under discussion, and whether its preliminary supply agreement converts into shipments.
- The next OECD count of venture capital into AI infrastructure and hosting, which is the cleanest test of whether the physical layer keeps its share.