# On Computers

*June 21, 2026 · By Keith Adams*

> A walk through the computer-brain industry (CPU, memory, accelerators) from a $15B footnote to a $350B heavy industry in 2025. Its recent, and we speculate, near-future doublings change where value accrues for infra founders and application builders alike.

Ever since my dad brought home an Apple IIe in 1984, I've loved
*computers*. Not computer science, or computer programming (though
eventually I'd come to love these too): the power-sucking, pixelated,
thrumming physical gadgets. I don't love them because they are
useful or intellectually stimulating, but because they are intrinsically
fun machines. I like it when they feep; when they chime at power
on; when their fans run, heating up the room; the ever-changing
acoustic and tactile feedback of their keyboards; and the clunking
and screeching and clicking of their storage devices, though there
is [less of that](https://www.amazon.com/dp/B00RXEWOAA) in these
solid-state days.  Computers are just *cool* to me, the way muscle
cars or mechanical watches are to their enthusiasts.

For most of my life, this fascination has been wholly disconnected
from the role these physical machines play in the economy. Even though
computers impact a bunch of important industries (software, IT services,
the Internet, etc.) the computers *themselves* have been kind of
marginal in capitalism's grand parade. It can be tricky
to decide what counts exactly as *computers* as distinct from semiconductors,
personal electronics, and other categories more cleanly captured in economic
statistics. For my purposes, I think of the "physical compute substrate" 
as being composed of CPUs, RAM, and accelerators where applicable (lately,
of course, GPUs). These three are often the principal components of what
problems a given computing device is fit to solve.

---

As the 2022 vintage AI boom has progressed, leading to ever more
financial heroics in datacenter construction, my childhood notion
of this "core" segment of the hardware market as a niche industry
has been feeling ... off. Financed with sophisticated combinations of equity
and debt, compute spend is driving multiyear funding plans that
have edged the margin-rich and asset-light hyperscalers of yore to
look more like classic heavy industries than the cottage industry
I came of age in. But I am in AI-besotted San Francisco near the
peak of a market cycle, and our intuitions can deceive us in moments
and places like this. What do the numbers show? And how will they
change, to the extent we can foresee?

Trying to guesstimate the revenue of a particular slice of an industry
like this always involves some guessing, but between
Epoch AI's epic chip sales dataset, WSTS's industry billings,
and a stack of Intel, Nvidia, and AMD 10-Ks,
we can put together a guess that computers *qua* computers were
about a $15B[^inflation] industry all-in in
1984[^methodology]. To give some sense of scale, that is
about the size of Major League Baseball. MLB is, of course, considerable,
and irreplaceable to its aficionados. But it is also not a lynchpin of
global commerce. I believe the US DoW has no contingency plans to wage war against
the Dominican Republic should it disrupt the baseball talent supply, for instance.

When I grew up and joined the workforce of software engineers in
2000, the computer brain industry had grown up to about $90B in
revenues; about the revenue of global distilled spirits.

2020, the last time I drew a paycheck as a software engineer, it
reached about $200 billion. About the global pet care industry.

But then in 2025, it nearly doubled to $350 billion. (Global cement.)
And we have every reason to expect this explosive growth to continue,
given the hyperscalers and the frontier labs' commitments to compute
spend over the foreseeable future.

| Year        | Revenue (2025 USD) | Biggest slice              | Comparable industry 
| :---------- | -----------------: | :------------------------- | :------------------
| 1984        |              $15B  | memory (DRAM)              | MLB
| 2000        |              $90B  | CPUs and memory            | Spirits
| 2020        |             $200B  | memory, CPUs, and GPUs     | Global pet care
| 2025        |             $350B  | datacenter AI accelerators | Cement
| 2026 (est.) |             $700B  | AI accelerators and HBM    | Cosmetics

The vibeshift around compute, then, is not wholly illusory. The economy
around the physical substrate for the computing revolution has spiralled up
in size from Major League Baseball to cement. If it should double again, as
seems very likely in the next year or two, it will be comparable to the
global cosmetics industry.  Another doubling from there would put it in the
heady air of the advertising and apparel industries, among the largest on
Earth.

---

## So what?

So, computing machinery is *bigger* than you think it is, and getting bigger
faster than you can update your intuitions about how big it is.
As technologists, we at Pebblebed find these machines fascinating. But we're also
investors, and with our fiduciary hats on, we're compelled to ask "[So
what?](https://now.fordham.edu/fordham-magazine/tribute-don-valentine-silicon-valley-pioneer/)"
OK, computers are blowing up. How do we act on this information?

### Infrastructure Software

One consequence of computing moving up an order of magnitude is
that *infrastructure software*, loosely defined as software that
exists to unlock latent capabilities in the hardware, can create
proportionately more value. When we were both at Facebook, my partner
Pamela Vagata once earned a coveted piece of corporate swag with
an understated patch casually boasting "$1B SAVED", with a graphic
suggesting the transition from exponential decay to exponential
growth. She saved the company over $1B by inventing the [ORC file
format](https://orc.apache.org/docs/), which radically improved
the storage and compute efficiency of many critical workflows.

Back in 2013 when she was doing this work, computing was too small to
support a venture-scale outcome for a company driven by these kinds of
insights. But in the context of 2026 hardware budgets, a comparable feat of
invention and technical derring-do might easily save $10B, or more.
"Savings" of this order of magnitude aren't best modeled as cost
reduction, but as unshackling the company to face enormously more ambitious
projects. Capturing even a small fraction of the value created in this way can 
lead to outcomes that would have been historic 10 years ago.

In the Pebblebed portfolio, we focus our infrastructure investing close to
the hardware/software interface. We believe that unlocking inefficiencies
and operational ease at this layer is going to explode over the decade
ahead. [Cedana](https://cedana.com/), which allows GPU jobs to be migrated,
increases neoclouds' revenue per MW. [Northflank](https://northflank.com/)
provides the undifferentiated heavy-lifting that has to happen to turn raw
k8s running on your cluster into a usable, observable, resilient system.
And [Lemurian Labs](https://www.lemurianlabs.com/) are building the modern
analog to the Java Virtual Machine, bringing write-once/run-anywhere to
accelerated compute.

### Application software

The boom in AI compute is radiating out into the larger economy in
various ways, and we are starting to see its impact in our application
software as well. [Build](https://build.inc) makes AI for the built
world.  In product terms, they build AI capable of performing
previously-labor-intensive information-gathering processes in commercial real
estate. From their start, Build has made hard choices to prioritize serving those
developers who are building datacenters. This was not an entirely
consensus perspective at the time of our investment a year ago, but
it has panned out even better than all but the most ardent bulls
would have predicted so far.


---

These are our perspectives to date. We don't have any glib conclusions, and
will not know with precision how this all turns out for a while.  Computing
has gotten much, much bigger, and sheer momentum guarantees that will
continue for a while. Anyone offering firmer conclusions than this is
probably epistemically overconfident. I feel grateful to my younger self
for finding these unusual machines so compelling.


[^inflation]: All dollar figures are inflation-adjusted to 2025 USD
equivalents unless otherwise noted.

[^methodology]: We are counting the merchant market for the
three things I'm calling a computer's brain: CPUs; memory
(DRAM, with a little SRAM); and compute accelerators crunching numbers beside them (the discrete FPUs of yore, and GPUs, inclusive of datacenter GPGPUs). With the help of my friendly [local AI](https://claude.ai), I have tried to tally the
worldwide spend in dollars for each chip, then used CPI to turn them into
2025 dollars. Recent accelerator figures lean on
[Epoch AI's chip-sales dataset](https://epoch.ai/data/ai-chip-sales). The rest is assembled year by year:

    * 1984:
        * CPUs: a [Dataquest worldwide table](https://archive.computerhistory.org/resources/access/text/2013/04/102723388-05-01-acc.pdf), checked against [Intel's 1984 annual report](https://www.intel.com/content/dam/www/central-libraries/us/en/documents/2025-05/history-1984-annual-report.pdf)
        * DRAM: the same [Dataquest table](https://archive.computerhistory.org/resources/access/text/2013/04/102723388-05-01-acc.pdf) (MOS-memory line), plus [USITC anti-dumping filings](https://www.usitc.gov/publications/701_731/pub1862.pdf)
    * 2000:
        * CPUs: [SIA](https://www.eetimes.com/sia-pegs-2000-chip-industry-growth-at-37-percent/) product-line billings, with [Intel's 10-K](https://www.sec.gov/Archives/edgar/data/50863/000091205701503434/0000912057-01-503434-index.htm) behind the figure
        * DRAM: [Gartner Dataquest](https://www.eetimes.com/dram-plunge-shuffles-top-10-chip-ranking-in-2001/), against the SIA billings
        * GPUs: [NVIDIA's 10-K](https://www.sec.gov/Archives/edgar/data/1045810/000101287001500492/0001012870-01-500492-index.htm)
    * 2020:
        * CPUs: [WSTS](https://www.wsts.org/67/Historical-Billings-Report) and [IC Insights](https://www.icinsights.com/news/bulletins/DRAM-Leads-In-Revenue-NAND-With-Top-Percentage-Growth-In-2020/), with the [Intel](https://www.sec.gov/Archives/edgar/data/50863/000005086321000010/0000050863-21-000010-index.htm) and [AMD](https://www.sec.gov/Archives/edgar/data/2488/000162828021001185/0001628280-21-001185-index.htm) 10-Ks behind the figure
        * DRAM: the same [WSTS](https://www.wsts.org/67/Historical-Billings-Report) and [IC Insights](https://www.icinsights.com/news/bulletins/DRAM-Leads-In-Revenue-NAND-With-Top-Percentage-Growth-In-2020/) lines
        * Accelerators: [Jon Peddie](https://www.jonpeddie.com/news/global-add-in-board-market-8-8-billion-in-q325-with-a-cagr-of-0-7-to-2029/) for the gaming boards; [Nvidia's segment filings](https://www.sec.gov/Archives/edgar/data/0001045810/000104581021000007/q4fy21pr.htm) for the datacenter
    * 2025:
        * CPUs and DRAM: the same houses as 2020, plus [TrendForce](https://www.trendforce.com/presscenter/news/20251218-12843.html) for the high-bandwidth memory
        * Accelerators: [Epoch AI](https://epoch.ai/data/ai-chip-sales)
    * 2026 (estimates):
        * Memory: [WSTS](https://www.wsts.org/esraCMS/extension/media/f/WST/7310/WSTS_FC-Release-2025_11.pdf) and [TrendForce](https://www.trendforce.com/presscenter/news/20260122-12893.html) projections
        * Accelerators: [Nvidia's quarterly run-rate](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000019/q4fy26pr.htm)

    The 1984 number has the spottiest sources. Back then memory, not logic, was
    the major revenue driver: in 1984 the world bought about $6.2 billion of memory chips but only $3.2
    billion of microcomponents (CPUs, and the microcontrollers and glue
    that ride along with them), of which true CPUs were a sliver. So,
    dollar-for-dollar, the brain of that Apple IIe was mostly RAM. (Those figures come off Dataquest's 1984
    worldwide table, scanned into the Computer History Museum's archive.)
    Also, an important judgement call: I am leaving out the processors in phones and tablets 
    (which I file under personal electronics, vs. computers), and I subtract
    the high-bandwidth memory soldered onto the AI accelerators from the memory
    column, since it is already counted in the price of the accelerator and
    I'd rather not double-count. Memory revenue is also violently
    cyclical (DRAM can halve in a year), so any single snapshot is partly a story
    about where we happened to catch the wave. 2026 is half a forecast,
    riding a memory supercycle; please do not take it as more than one
    significant digit.

---

**Keywords:** computing, semiconductors, AI, datacenters, investing