AI Capex & the 1999 Telecom Cycle
What the Build-Out Rhymes With, and What It Does Not

The mainstream view in 2026 is not that AI is cheap or that risk is absent. It is that the present cycle is structurally different from 1999 in five specific ways, and each carries weight.

One, profitable incumbents, not speculative startups. The 2026 capex is being funded by Microsoft, Alphabet, Amazon, and Meta. Trillion-dollar businesses with operating cash flows that would have been unimaginable to a 1999 carrier. The buildout is internally financed off a foundation of recurring software and advertising revenue, not from junk-rated bonds underwritten by hopeful equity holders.

Two, prior capex is visibly converting to revenue. Q1 2026 earnings showed cloud margins expanding, Meta and Google advertising re-accelerating to multi-year highs, AWS at 28 percent year-over-year growth, Microsoft’s contracted backlog up 99 percent. The argument is that the 2023–2024 cohort of AI spend is already in the P&L, and the 2025–2026 cohort will follow.

Three, valuation, while elevated, is not 1999. Today’s forward P/E on the S&P 500 sits at roughly 21×, expensive but a long way from the 28× trailing multiple of late 1999. The Magnificent Seven trade at 30 to 35× forward earnings on average; Cisco at the March 2000 peak traded above 150×. By the headline-multiple measure, today is not 1999.

Four, demand is here, not in 2010. The fiber laid in the late 1990s was about a decade ahead of its profitable end-market. ChatGPT has 900 million weekly users today. Enterprise AI deployments are live and billing. The use cases are not speculative; they exist.

Five, the financing is internal, not vendor-extended to fragile counterparties. Cisco, Lucent, and Nortel lent billions to CLECs and long-haul carriers that could not have paid cash. Today’s hyperscalers are paying Nvidia in cash. The closed-loop financing risk that characterized 1999 is, on this view, structurally absent.

The shape of passive exposure has changed. A decade ago, owning the S&P 500 meant owning a broad basket of 500 American businesses where the top ten were ~19% of the basket. Today the top ten are ~41% of it. The remaining 490 names collectively carry the minority weight. When you buy SPY today you are buying a concentrated position in roughly ten companies, most of them in technology and AI, supplemented by 490 names that increasingly do not move the dial. The disclosure label has not changed; the underlying risk profile has, substantially.

Why this matters before the Cisco discussion. The 1999 cycle was carried by telecom equipment makers and carriers; they were a large weight but not 41 percent of the index. Cisco at its 2000 peak was the largest single stock at roughly 4 percent of the S&P 500. Today’s leadership names are individually larger, and collectively much larger, than 2000’s. That changes the math on what a Cisco-style outcome in the AI cohort would do to the index, and to anything passively benchmarked against it. The next section on valuation, and the Cisco overlay that follows, both need to be read with this context in front of them.

The headline forward P/E at ~21× sits above its 5- and 10-year averages, but headlines are flattered when corporate earnings sit at peak margins, as today’s do (the aggregate S&P 500 net margin in Q1 2026 hit a series-high 13.4 percent). CAPE was designed precisely to strip that flattery out. At ~42×, today’s reading is more than double the 140-year average and within two points of the December 1999 record. A published estimate this month put the implied forward annual real return at roughly 1.3 percent. That is not a forecast I make; it is what the historical regression says.

The “profitable incumbents are different” defense is the most commonly cited point and the one I want to address first, because I have firsthand recollection of the 10-Ks. WorldCom was the second-largest US telecom by revenue, capitalized at over $115 billion at peak. Sprint, AT&T, and the Baby Bells were not speculative ventures. They were the institutional holdings in every diversified pension portfolio.

The Cisco data point is the one I keep coming back to. Cisco was Nvidia. It sold the picks and shovels to a real, growing, profitable industry. At the March 2000 peak it had grown net sales 53 percent year over year, was the largest company in the S&P 500 by market capitalization, was profitable to the tune of $2bn in net income, and was trading at 150× forward earnings. It then fell 85 percent over the next two and a half years, and took more than 25 years to regain that high, which it passed only in December 2025. Cisco never went bankrupt. The malinvestment by its customers showed up in Cisco’s own equity. Profitable incumbency did not protect its shareholders.

Now place Nvidia on this chart. Nvidia is selling the most important shovels of the current cycle, just as Cisco did in 1999. Nvidia’s customer concentration is real: per its own 10-Q, hyperscalers were ~50 percent of data-center revenue in Q1 FY2027, and revenue is “concentrated among a limited number of direct and indirect customers.” If those customers ever cut capital programs in unison, Nvidia’s revenue line goes with them, regardless of how excellent the underlying technology is. That is not a forecast. It is what the customer composition tells you about the shape of the risk.

The natural counter is that Nvidia’s forward P/E is ~50×, not Cisco’s 150×. That is correct. It also matters less than the headline suggests. A 50× multiple still requires roughly a tripling of earnings to grow into a 15× “normal” multiple, and that tripling has to occur while customer capex stays elevated and unit pricing holds. Both are exactly the conditions in question.

The bull-case claim that “today’s hyperscalers are paying Nvidia in cash” is half-true. The hyperscalers themselves are. But the ecosystem they depend on for end-demand on their AI services is not, and the most aggressive growth in compute purchasing is happening at OpenAI and the neoclouds where the cash is coming from Nvidia or from debt secured against future contracts.

I would frame this conservatively. The 2026 financing structure is less aggressive than 1999. The vendors are bigger and have more cash. But the closed-loop element is not absent, and where it exists it sits at the most fragile end of the demand stack, the AI model companies and neoclouds whose own end-customer economics are exactly the open question of the cycle.

The story the three charts tell in sequence. (4a) Capex has roughly tripled in two years, from $230bn to a projected $725bn for 2026. Each earnings round so far has revised this number upward. (4b) That spending now absorbs close to 100 percent of operating cash flow for the AI cohort, against a ten-year norm closer to 40 percent. The cash that used to be available for shareholder returns, debt reduction, or strategic acquisition is now being consumed by the build. (4c) The cohort has split. Alphabet bought back zero stock in Q1 2026, down from $15.1bn a year earlier, despite holding $127bn in cash. Meta did the same. Apple, which deliberately sat out the AI arms race, did the opposite, running a second straight $100bn buyback against only ~$13bn of full-year capex. The divergence inside the Magnificent Seven is now as important as the group itself.

What this looks like to a long-only allocator. The companies in question have not stopped being profitable. They have stopped returning capital. For an equity holder, the present value of expected future buybacks and dividends has been redirected toward capex that may or may not clear its cost of capital. If it does, that capital allocation will eventually be vindicated and shareholder returns resume. If it does not, the forgone capital return is the first loss; the second is the writedown on the capex itself. The shareholder is implicitly bearing both risks for some indeterminate period.

This is the part of the analogy the 1999 generation did not have to deal with as severely. Long-haul fiber commoditized over four years, from about 2000 to 2004. AI inference is commoditizing in real time, with frontier models becoming “obsolete within three weeks of launch” per one widely cited analysis. The capital being deployed against AI demand is being deployed against a product whose unit price is collapsing at 75 percent per year.

For the capex math to work, end-revenue from AI services has to outrun two exponentials at once: rising end-demand AND collapsing price per unit of compute. In the 1990s telecom cycle, only the unit-price exponential mattered. The demand exponential held for another decade. Today both clocks are running, and they are running against the same buildout.

Each sleeve runs continuously and is sized within bounds set by the macro overlay. The combined book targets near-zero beta to the S&P 500 and earns its spread from the relative-value inefficiencies in the listed US equity complex. Capacity is a function of instrument depth rather than the smallest position; the sleeve venues above are deliberately deep-liquidity for that reason.

The net read. The bull case is not wrong. It is just less protective than its proponents argue. The five differences from 1999 give the cycle a roughly 30 to 50 percent margin of safety on starting valuation, balance-sheet quality, and demand reality. They do not give it immunity. And the new element of the 2026 cycle, the rate at which the underlying product is commoditizing, has no equivalent in the 1999 setup, and runs in the opposite direction of the bull case.

I am deliberately not building these into a recession call. Global consumer confidence actually firmed slightly in May 2026 (Ipsos 47.4, up 0.7 month on month), and US jobs remain resilient. The point is narrower and worth stating cleanly: the financing and demand backdrop is not supportive of an unprecedented, debt-funded capital programme, and a market priced for perfection has little room for the consumer or the cost of money to disappoint.

Two specific signals on the rate side deserve attention. First, the US 10-year sits near 4.6 percent, close to the top of its 12-month range, after 75 basis points of Fed cuts since September 2025. The Fed cutting has not eased long-term financing costs, which is itself information. Second, the UK 30-year gilt at 5.78 percent is a 1998 high, and Alphabet has already sold a 100-year sterling tranche this year to fund the build. The rate environment is not just unsupportive at the front end; it is repricing long-duration debt upward globally exactly as the AI cohort is term-funding a multi-year capex commitment. I write that having spent the last two decades watching long-end pricing in EM and DM debt markets at close range; the speed of the global repricing this cycle is the most consistent thing across them.

On the consumer side, the divergence is the unusual fact. The Conference Board index sits at 93.1; two of three Americans report cutting back spending even as the market hits new highs. Healthy markets and gloomy households can coexist for a while, particularly when employment holds. The point is that the gap is wide and the explanation for it (energy prices, food, lingering tariff uncertainty) is not the kind that resolves quickly. If the labour market wobbles, the divergence closes downward, not upward.

Connecting it to the cycle thesis. A debt-financed capital programme of unprecedented scale is being underwritten against an exponentially growing demand forecast, by companies whose end-customers are increasingly cautious about discretionary spending, into the highest long-term borrowing costs in a generation. None of those four facts in isolation overturns the bull case. The aggregate increases the variance of the outcome materially, and that is what positioning has to respect.

I will not pretend to forecast a top. I read tops worse than I read bottoms, which is itself a lesson from being on the Salomon floor in 2000 and not being on the Greenwich floor in 2009. I do not need to call the top. I need to position around the shape of the risk, and the shape is now well-described enough to act on.

Concentration of outcome. The S&P 500 is now an unusually levered bet on the success of an AI capex cycle being underwritten by a small cohort. Passive US equity exposure is, in 2026, less diversified than it has been in living memory. The SPY ETF carries the same disclosure label as ever; its underlying risk profile has changed substantially.

What Britannica does. The fund is systematic, market-neutral, US equities. It does not bet on the direction of the index. It is constructed to earn the spread between long and short books with broad market direction hedged out. When the index leans on a few correlated names making the same bet, capital that does not depend on that bet is structurally valuable. We are positioned around the shape of the risk, not a prediction of timing. If the AI cycle clears every test in this note, the book still earns its spread. If it does not, the book is not a passenger.

What is genuinely true on the other side, and worth respecting. Durable software earnings really are different from dot-com fiber as an asset class; the underlying economics have a backbone that 1999 did not. And the discipline of knowing companies deeply, which is what the best long-tenured managers do, is exactly the kind of discipline that can navigate a regime where flow dominates the index but does not dominate every security inside it. The bear case here is about passive US equity exposure and the AI-cohort capex math. It is not a claim that fundamental work has stopped working at the security level.

What would change my mind. If hyperscaler capex guidance slows materially at the late-July prints. If AI unit pricing stabilizes and growth catches up. If the top-10 weight begins normalizing rather than rising further. If the OpenAI usage trajectory re-accelerates against credible competitors. Any of these would walk back the position one notch. None of them have happened yet.