Musk as a Service
Every position in the AI trade can be individually right. The trade in aggregate cannot.
Author's note: A companion piece to "The Golem in the Server Rack." Golem argued the silicon underneath the AI buildout will commoditise faster than the buildout assumes. This piece asks the next question — what happens to the equity positions priced for that buildout earning everything its modelling implies, and which break first when the math doesn't close. The cleanest entry point is the most aggressively AI-maximalist position in public markets: the integrated Musk portfolio, which we should probably just start calling Musk as a Service.
There is a way to read the AI trade that nobody quite says out loud, because saying it out loud makes the rest of the trade look unsustainable.
The way is this. The market has assembled a portfolio of bets on the AI revolution that are each, individually, defensible. The Tesla bull is making a coherent bet on long-tail product capture. The NVIDIA bull is making a coherent bet on near-term picks-and-shovels dominance. The hyperscaler bull is making a coherent bet on monetising the buildout through enterprise software. The MaaS bull is making a coherent bet on Musk personally delivering the full physical-AI future. Each story can be told with a straight face. Each multiple can be defended on its own theory.
The problem is what happens when you add them together. The aggregate price tag of every position currently positioned to win from the AI revolution exceeds what that revolution can plausibly produce in present-value cash flows even if each bet is individually correct. That's not a story about someone being catastrophically wrong. It's a story about everyone being a little bit right and the math still not closing.
That's the dark-fibre pattern: individually rational bets whose aggregate capacity exceeded what the output price could sustain, even though demand grew exactly as predicted. It's the analogy this piece is built on — and I'll come back to it in full.
Doug O'Laughlin's October 2023 essay on the telecom bubble1 laid out this analogy at the industry-cycle level — capital intensity, leverage, demand-projection statistics, supply reacting to demand. This piece picks up where the macro framing ends: not whether the cycle resolves, but which positions inside the trade break first, and why the shape of the resolution will differ from the dark-fibre case in specific ways.
The Integrated Position
If you want to find the maximum-bullish-on-AI-revolution position currently being expressed in the public markets, you don't look at Tesla, or SpaceX, or xAI individually. You look at what we should probably just start calling Musk as a Service — the integrated portfolio of Tesla, SpaceX (which now includes xAI and X following the February 2026 merger), Optimus, Neuralink, and the Boring Company — valued in aggregate at somewhere approaching $3.5 trillion as of mid-2026, depending on which day you mark Tesla and how you value SpaceX ahead of its June 12, 2026 listing. The dollar weighting is dominated by Tesla and SpaceX; Neuralink and Boring are conceptually part of the portfolio but rounding errors in the headline number. What matters isn't the precise dollar total — it's that the market treats these positions as correlated bets on the same underlying thesis.
Tesla's roughly 215x forward earnings is the most visible piece.2 The car business does not justify the multiple, and Tesla bulls don't pretend it does. Tesla is priced as an AI-and-robotics company — autonomous vehicles producing software-margin annuities, Optimus humanoid robots at scale, FSD as a subscription, Dojo as a compute platform in its own right.
But Tesla on its own undersells the bet. SpaceX, post its February 2026 all-stock acquisition of xAI, is the more striking piece. The merger priced the combined entity at $1.25 trillion — SpaceX at $1 trillion, xAI at $250 billion. SpaceX is preparing to list on June 12, 2026, at a targeted valuation as high as ~$1.75 trillion — roughly 95 times sales for a company that lost money last year, and what would be the largest public listing in history.3 The merger isn't incidental detail — it's structural validation of the MaaS framing itself. Until February, treating Tesla, SpaceX, and xAI as one bet was a rhetorical convenience. Now two of the three are literally one corporate entity, with X folded in for good measure. Starlink as the global mesh underneath orbital data centres, the AI Sat Mini constellation I covered in the Golem piece, xAI as the model layer, X as the distribution layer, all inside one cap table.
Optimus is the robotics layer. Neuralink is the brain-computer-interface layer for whenever it's ready. The Boring Company is the urban infrastructure layer. Each of these is priced on the assumption that the AI revolution arrives in a specific maximalist form — autonomous everything, robotic everything, ubiquitous compute, distributed intelligence, the entire cyberpunk-but-helpful future. And each is priced additionally on the assumption that Musk personally is the one delivering it, because the brand, the capital-raising machinery, and the talent flywheel are centralised on him.
That's why the whole thing reads best as a single platform. The companies are legally separate, but the capital flowing into them isn't really pricing them as separate businesses — it's pricing them as correlated stages of the same single bet. Tesla shareholders trade Tesla on SpaceX news. xAI valuation rounds move Tesla. Optimus demos move Starlink. The "Musk premium" or "Musk discount" is a recognised market phenomenon — analysts have to model it. So MaaS exists in the way markets actually function, even when it doesn't exist on a 10-K. The merger just made it structurally real as well.
If MaaS is priced correctly, the AI revolution is happening in roughly the form Musk and his bulls describe — trillions of dollars in new economic value created by autonomous transport, humanoid labour, satellite-based compute, and consumer-grade artificial intelligence at scale. A hardware-and-software revolution comparable to the transistor or the internet, arriving inside a single decade. It is, by some distance, the most aggressively bullish AI-arrival thesis you can find in the public markets.
The Picks-and-Shovels Seller, Priced for Continued Acceleration
Now look at NVIDIA. NVIDIA is the company that builds the silicon the entire revolution MaaS is priced for actually runs on. It trades at roughly 22-26x forward earnings, give or take by the quarter.
That multiple is harder to dunk on than I would have said a few months ago. NVIDIA's most recent quarter posted revenue up 85% year-over-year, with the following quarter guided to roughly $91 billion in revenue.4 Gross margins are holding at 75%, the top of the recent range. At those growth rates, a forward multiple in the low twenties isn't a market signalling impending displacement — it's a market struggling to keep the forward-earnings denominator updated fast enough to catch up with reality.
But that's exactly where the position is more exposed than it looks. The 22-26x multiple is priced for continued explosive growth — Jensen's pitched "trillion dollars of Blackwell and Rubin revenue" thesis requires the hyperscalers to keep ordering at the current pace through 2027 and beyond. NVIDIA's current earnings are bulletproof. NVIDIA's forward earnings depend on the hyperscalers continuing to buy at scale through the next two product cycles. If the hyperscalers slow — if impairments start, if depreciation extensions get reversed, if the buildout hits the same kind of "rationalisation" moment that telecom did in 2002 — NVIDIA doesn't lose money on the hardware they already shipped, but the forward multiple gets cut hard.
So NVIDIA at 22-26x isn't, on its own, evidence that the market has secretly priced in disruption. The simpler reading is that NVIDIA is priced exactly as you would expect a company whose earnings are accelerating into a generational infrastructure cycle. The bet is internally coherent. The vulnerability is that "internally coherent at this multiple" requires the cycle to keep accelerating — which is the same assumption every other position in the trade is also requiring, and which collectively they can't all be paid for.
The Hyperscalers, Priced Like Software Companies
Now look at the companies spending the trillion dollars.
Microsoft trades around 22-25x forward earnings. Alphabet around 23-27x. Meta around 18-19x. These are multiples consistent with mature, dominant software businesses growing at respectable but unspectacular rates. Importantly, they are essentially in the same range as NVIDIA, which is worth pausing on — the data does not actually show "the market preferring hyperscaler equity over NVIDIA equity per dollar of forward earnings." The data shows them being priced at very similar multiples, which is what you would expect if the market believed they are all reasonably-positioned bets on AI growth, captured at different points in the stack.
What the data does show, more subtly, is that these multiples are not consistent with the hyperscalers being the primary beneficiaries of a generational AI revolution committing $80-145 billion per company per year of CapEx. If Copilot were going to be the next Windows, if Gemini were going to be the next Search, if Bedrock were going to be the next AWS at margin levels comparable to the originals, the hyperscaler multiples would have re-rated upward dramatically over the last two years. They have re-rated some — but nothing close to what you would see if the equity market actually believed the buildout would earn its capital cost back at the IRRs the CFOs are modelling internally.
So the hyperscalers' multiples are internally coherent — modest software-business multiples — but they don't separately underwrite the AI CapEx as a stand-alone thesis. The market is pricing these stocks for the software business that already exists, with the AI CapEx as a feature rather than the thesis.
Late-1990s telecom carriers had the cash flow to fund the buildout, the boards approved it, the technology worked exactly as advertised — and the equity market, by 2002, had repriced every single one of them on the realisation that the assumed returns weren't there. The hyperscalers are richer, better-diversified, and unlikely to actually go bankrupt. But the equity-market discount of the CapEx thesis is similar in shape, if not in degree.
The Coherent Reading, and Why It Still Doesn't Close
Put those three pieces next to each other and the obvious question is: are these positions even contradictory? Because there's a perfectly coherent way to hold all of them at once, and it doesn't require any of them to be wrong.
A portfolio constructor can rationally hold NVIDIA + Microsoft + MaaS as a sequential value-capture bet on the same revolution:
NVIDIA captures the next three to five years — the buildout phase. CUDA is the only credible production toolchain, and the hyperscalers can't yet refresh hardware fast enough to displace it.
The hyperscalers capture the following three to seven years — the monetisation phase. Copilot, Gemini, Bedrock, and their successors turn the buildout into enterprise software revenue at scale.
MaaS captures the long tail, ten years and out — the products phase. Cheap AI capability flows through into autonomous vehicles, robots, satellite networks, and consumer-facing software annuities.
That isn't three contradictory bets. It's one bet on the AI revolution, split across three value-capture windows and allocated to the equities most likely to dominate each. The same investor holds all three because they're underwriting the whole arc.
Read this way the multiples aren't incoherent at all — but it's worth being precise about why. A multiple isn't a measure of how long the value-capture window is; it's a measure of how big a company's future earnings are expected to be relative to its current earnings. NVIDIA at 22-26x reflects huge current earnings already in the numbers, with three-to-five more years of buildout dominance on top. The hyperscalers at 18-27x reflect enormous current cloud and software earnings, with AI as meaningful but incremental uplift on a base already at scale. Tesla at ~215x reflects tiny current earnings against a market cap pricing in the future birth of a robotaxi-robotics-satellite-AI business that barely exists today. Each multiple is roughly right for the ratio between what the company earns now and what the market expects later — and for these three, that ratio happens to line up with sequential phases of the AI cycle.
So why is it still a problem? Because individually coherent bets can still be collectively overbuilt — and a good deal of what looks like incoherence isn't even active belief.
Start with the part that is belief. Look at the aggregate: Tesla at ~$1.6T, SpaceX post-merger at ~$1.25T (targeting ~$1.75T at its June 12, 2026 IPO), NVIDIA at ~$4.4T, the four major hyperscalers at a combined $13-15T, Oracle at ~$700B substantially riding the OpenAI counterparty trade, plus several hundred billion in AI-startup private markets — roughly $20-22T of equity priced for the buildout earning everything its modelling implies. Add the second-order positions — Broadcom, AMD, TSMC, the broader supply chain — and the visible AI-trade aggregate runs north of $25 trillion.
Now the part that isn't. A large slice of that figure is mechanical exposure, not conviction. Vanguard owns Tesla and NVIDIA and Microsoft and Meta because its indices require it, not because its portfolio managers believe every thesis. The overbuilt observation survives the point — but it's moderated by it. A chunk of "the market" isn't holding contradictory beliefs so much as contradictory tickers.
For the conviction slice to earn the returns its multiples imply, the AI revolution has to deliver, in present-value cash-flow terms, something on the order of a global economy's worth of new enterprise value within fifteen years — net present value comparable to the entire current US GDP, with the cash flows accruing to the public-market equities currently priced for it. Not "AI is transformative." Not "AI will be huge." That specific.
It can happen. It's not the base case in any serious model. Some value leaks to private companies that aren't priced. Some leaks to consumers as surplus. Some fails to materialise because cost curves compress prices faster than demand expands. Some accrues to companies that don't yet exist.
There's a historical resonance worth pausing on. In the late 1990s, WorldCom's claim that internet traffic was doubling every 100 days drove the bull spirit of the entire telecom buildout — repeated by the government, the press, and the analyst community, and eventually wildly inaccurate (actual growth was closer to doubling every year). In 2018, OpenAI published that AI training compute had been doubling every 3.4 months since 2012. The two numbers are nearly identical: 100 days versus roughly 103.5 The rhetorical role is the same — justifying capital intensity at unprecedented scale by extrapolating a near-vertical curve. The 1990s statistic underwrote something on the order of $200 billion of telecom CapEx in today's dollars. The 2018 statistic and its successors are underwriting an order of magnitude more.
The individual multiples are defensible. The aggregate is not.
The Asymmetric Discipline
But here's where the coherent reading begins to crack in real-time. The hyperscalers are being penalised for their AI CapEx — specifically, for the merchant-GPU portion of it. Meta dropped 11% on its Q3 2025 earnings — its worst single day in three years — after raising capex guidance, then took another 6-7% hit in April 2026 when 2026 capex was guided to $125-145B.6 Microsoft has faced increasing analyst scrutiny on its capital-to-free-cash-flow trajectory. The same April 2026 print that sank Meta saw Alphabet and Amazon — also reporting heavy AI CapEx — rise on the day because they showed revenue translation and visible in-house silicon traction. The market isn't punishing AI CapEx in the abstract. It's punishing CapEx that doesn't show movement away from NVIDIA dependency.
Musk's portfolio is doing structurally the same thing. xAI's Memphis supercluster is one of the largest NVIDIA deployments outside the hyperscaler set. Starlink's orbital programme carries substantial merchant-silicon procurement for its ground stations and node infrastructure. The market does not punish them. Tesla's multiple has expanded through the same period in which the hyperscalers' has compressed. SpaceX's valuation rises with every funding round. xAI has raised at increasing valuations into the headwind that's dragging the public AI-CapEx names.
The multivariate caveat is worth naming. Tesla's multiple expansion isn't only a narrative story — FSD milestones, Optimus demos, robotaxi unveils, China-delivery numbers all do work too. The point isn't that narrative is the only driver of the asymmetry. It's that even setting the operational milestones aside, the GPU-CapEx component of Musk's spend receives discipline on one side of the trade and approval on the other. The structural exposure is the same. The market reaction is asymmetric.
The asymmetry is structurally interesting because the engineering bets underneath are almost identical. Microsoft, Google, and Meta are running in-house silicon programmes — Maia, TPU, MTIA — designed to substitute away from NVIDIA's gross margin. Musk's orbital programme is also ASIC-led, built around transformer-optimised inference silicon rather than GPU racks. Both sides have priced in the same silicon trajectory. The deeper mechanism the market has internalised but doesn't quite name is a transition-speed metric. NVIDIA spend reads as failure-to-transition. In-house silicon spend reads as the disciplined response. Google has never depended on NVIDIA for its own AI workloads — Gemini runs entirely on TPU — the existence proof that you don't need NVIDIA if you've built the alternative. Microsoft, AWS, and Meta are visibly racing to catch up, and the market judges them on the trajectory of in-house adoption, not just on absolute CapEx. Musk's portfolio escapes the same framework because there's no captive in-house silicon programme the market is grading against. When xAI buys NVIDIA at scale, the market reads it as acquiring necessary growth infrastructure, not as failure-to-transition. Same dollars, different framing, different reaction.
The asymmetry isn't about what's being built. It's about what gets to be a story. The hyperscalers can't sell their silicon transition at a premium multiple because "we are doing the same thing we have been doing for five years, just more of it" doesn't move equity markets. Musk can sell the same structural bet as space colonisation because the narrative has enough cinematic texture to sustain the premium. Same silicon, different theatre. The buildout-discipline curve has already arrived for one side and not the other, and the gap is the gap between engineering substitution and spectacle. The question this piece is built around is when the market closes that gap on the other side.
Why the Race to the Bottom Wins
The mechanism is fundamental economics, not Moore's Law. In a market with multiple suppliers, falling input costs, and an undifferentiated output — and a token is a substantially undifferentiated output, regardless of which company served it — competition drives output prices toward the marginal cost of the cheapest serving infrastructure.
The Golem piece makes this case at length; the compressed version is that the dark-fibre comparison is imperfect in one specific way. AI compute isn't one market but a set of stack-bound ones — NVIDIA/CUDA, AWS/Trainium, Google/TPU, Microsoft/Maia — each with its own toolchain and developer base. Buyers aren't trapped inside a stack so much as anchored to it; leaving costs real integration work. So the race to the bottom runs fast inside each stack (newer silicon undercuts older every eighteen months) but moves on a generational timeline across stacks. That doesn't dissolve the gravity. It does mean the buildout looks profitable inside each stack for years longer than the telecom comparison alone would predict.
Even with that friction, the evidence is already on the tape. GPT-4-class capability cost roughly $30 per million tokens in 2023; equivalent capability in mid-2026 is closer to $0.50 to $2, depending on provider and routing — a >95% collapse in two and a half years while infrastructure CapEx multiplied. Epoch AI's benchmark-controlled analysis7 puts the decline to match GPT-4-class performance at roughly 40x per year, with the steepest drops arriving after 2024.
The hyperscalers are racing each other to that floor. Microsoft uses Maia to undercut AWS, AWS uses Trainium to undercut Microsoft, Google uses TPUs to undercut both. The cost advantage each captures from its own silicon isn't margin it keeps — it's the cudgel it hits the others with. NVIDIA-priced capacity has to compete at the floor the in-house silicon sets.
The depreciation move makes the squeeze worse. Across 2023-2025 the major buyers mostly stretched server useful-life assumptions toward five or six years — an accounting choice that flatters current earnings by spreading the same cost over more years, and a bet that the hardware stays economically useful that long. Against an eighteen-month NVIDIA cadence (Blackwell → Blackwell Ultra → Rubin), each generation less competitive on watts per token than the last, the gap between accounting life and economic life is widening — and Golem details why.
The Pattern, Cycle by Cycle
This shape is not new. The crucial detail that matters for the AI cycle is that each of these prior cycles featured individually rational bets that were collectively overbuilt — not investors holding contradictory positions, but coherent positions whose sum exceeded what the underlying economics could pay for.
Telecommunications dark fibre, 1998-2003. Multiple carriers — WorldCom, Global Crossing, Level 3, Qwest, 360networks — laid millions of miles of fibre on the assumption that internet traffic growth would absorb it. Each carrier's bet was internally rational: traffic was growing, each carrier had a plausible regional or strategic angle. The traffic growth was also real — internet traffic did grow as predicted. The carriers nearly all went bankrupt anyway, because the aggregate capacity they built exceeded what the price-per-bit could sustain. The fibre is still in the ground. Most of it spent a decade as "dark." The technology won. The capital that paid for the first wave was destroyed.
Bitcoin mining ASICs, 2017-2022 — the highest-resolution analogue available, because it's the only one of the three cycles where silicon obsolescence has played out continuously and at sub-annual cadence against a visible difficulty metric. Each generation of mining silicon — Antminer S9, S17, S19, S19 XP8 — made the prior generation unprofitable as network difficulty rose. Each individual mining company was internally rational, betting on its current-generation fleet at the time it deployed. The aggregate of all mining companies' deployments raised difficulty faster than any single company's fleet could amortise. Asset values collapsed not over years but over months. The companies that won were the ones that could refresh hardware fastest, not the ones with the biggest existing fleets. What made the cycle move so fast was that obsolescence tracked one-to-one with power efficiency — the S19's electricity bill stopped balancing against block-reward share the moment the S19 XP improved joules-per-terahash. The AI equivalent is joules-per-token-of-inference, which is the metric every ASIC programme racing against NVIDIA is now optimising for. This is also the cycle the AI buildout's actual operators lived through up close. Tesla held BTC on its balance sheet in 2021. Altman launched Worldcoin in 2023. The operators running the most consequential AI buildout have direct experience of how electricity-to-digital-output economics resolve, and they're constructing the AI compute market with that experience baked in.
Solar panel manufacturing, 2009-2012. Module prices fell from roughly $4 per watt to roughly $1 per watt over four years9. Each manufacturer was internally rational, building capacity at the prices they could see when they broke ground. Aggregate capacity ran ahead of demand growth, and prices fell faster than capital deployed at the higher prices could amortise. Bankruptcies cascaded — Solyndra10 is famous, but it had plenty of company. The technology won, decisively. The first wave of capital deployed to build it did not.
The common shape: improving technology in a competitive output market drives output prices toward marginal cost faster than the deployed capital can amortise, even when every individual deployment was rationally underwritten on the prices visible at the time. The technology wins in every case. The capital that paid for the first wave gets destroyed in every case. Not because anyone was stupid. Because everyone was right and the sum was still too big.
Who Survives the Curve
The interesting question isn't whether the trillion dollars gets repriced. It's who's positioned to survive the repricing.
Start with NVIDIA, the picks-and-shovels seller. It gets paid up front, in cash, for hardware the buyer carries on its balance sheet — so its exposure is to demand (does the buyer keep buying?), not to ROI (does the buyer earn its capital cost back?). That's structurally better than the buyers' position.
This is what is currently showing up in the financial reports. NVIDIA's gross margins are holding at roughly 75%, the top of the recent range, with quarterly revenue growth running at 85% year-over-year. Microsoft's and Meta's free cash flow growth has slowed sharply as CapEx has eaten the operating leverage. Oracle has taken on enormous counterparty exposure to OpenAI's ability to actually generate the revenue to pay for the compute it has contracted. None of this fits the narrative of "the trillion dollars is paying off broadly." It fits the narrative of "the chip seller is in the strongest position; the chip buyers are stretched."
But "structurally better" isn't "invulnerable": the same multiple is priced for continued explosive growth, so a normalisation in order pace could cut the stock meaningfully even without a fundamental break in the business. Better than the buyers' position; not riskless.
Musk as a Service, in this framing, is a different bet entirely — a bet that the AI revolution flows through to products rather than infrastructure. If the MaaS bulls are right, the economic value gets created downstream from the chips and the datacentres, in the cars and robots and satellite networks and software annuities that use the cheap AI capability the buildout produces but can't profitably charge for. The MaaS bet and the hyperscaler bet aren't strictly contradictory — both can be true on different timelines — but they have very different sensitivities to the cost curve. Cheap AI inference is good for MaaS (lower input costs for the downstream products) and bad for the hyperscalers (lower output prices for the services they sell).
The Question Everyone Is Pretending Not to Ask
Every overbuilt infrastructure cycle in history eventually forces the aggregate to resolve. The dark-fibre carriers had to admit by 2002 that the traffic growth wasn't going to pay for the build. The solar manufacturers had to admit by 2012 that module prices weren't going to support the factories. The Bitcoin miners learn it every cycle on a shorter timeline.
One structural difference from the 2000 case is worth naming. The dark-fibre carriers were highly levered — WorldCom alone carried roughly $30 billion of debt against a peak market cap of $180 billion; the smaller carriers were debt-financed end to end. When the cycle turned, the resolution was bankruptcy: equity zeroed, debt restructured, businesses absorbed. The hyperscalers in the AI cycle are the opposite. Microsoft, Google, Meta, and Amazon collectively hold net cash, not net debt. This cycle won't resolve through insolvency. It will resolve through earnings impairment, multiple compression, and write-downs — pain at the equity level, not at the business level. The shape of the resolution differs from 2000. The aggregate of equity value destroyed could still be comparable.
The AI buildout will resolve in the same way, on some timeline. The question that nobody in the investing class is comfortable asking out loud is: which position gets written down first, and how big does that write-down have to be before the others reprice in sympathy?
My guess — and I want to flag it as a guess — is that the first major impairment shows up not as a clean write-down but as a quiet utilisation footnote in a hyperscaler 10-K, sometime in 2027 or 2028. The depreciation-extension move has already been used — and is still being reached for, with at least one hyperscaler stretching server lives again into 2026 — but each turn of it buys less and gets harder to disguise. The official framing will be that the chips were "fully utilised" but that "demand mix shifted" toward newer hardware. The implicit framing will be that the bet was the wrong scale, even if the underlying technology assumptions were broadly correct.
In that scenario NVIDIA survives as a business but not as a multiple — the explosive-growth premium gets cut when the explosive growth normalises. The hyperscalers' shareholders are the ones most exposed, holding stocks at multiples that already don't underwrite the CapEx but do underwrite a software business now competing on AI-token prices that are 95% lower than they were two years ago. MaaS is exposed on a different vector: it needs cheap AI inference to flow through into product economics, which it will get, but it also needs the autonomous, robotic, satellite-connected version of that future to actually materialise on the timeline the multiple is pricing.
The AI revolution will happen, more or less on the silicon trajectory the Golem piece lays out. Musk as a Service may or may not play out as the maximalist version of the downstream story it's priced for.
The piece's thesis, stated plainly: the aggregate trade is structurally overbuilt, and resolution arrives sometime between 2027 and 2029. The hyperscaler-CapEx side cracks first — most likely via a quiet utilisation footnote in a 10-K, framed as a demand-mix shift rather than an impairment. The Musk-narrative side follows on its own timeline, when the orbital and robotaxi stories begin to be priced against actual delivery rather than against vision. What any of this means for any individual position depends on cost basis, time horizon, and tolerance for being early. The argument is structural. The timing is the variable nobody can call cleanly.
Every position in the AI trade can be individually right. The trade in aggregate cannot. That's the dark-fibre pattern. It is also, much less famously, the pattern of basically every major infrastructure cycle in capitalist history. The technology won. The first capital deployed to build it did not.
You can be a fully-committed AI bull — even a fully-committed MaaS bull — and still believe that. In fact, if you've held the position long enough, you probably already do.
Doug O’Laughlin’s October 2023 essay lays out the telecom-bubble analogy at the industry-cycle level — and corroborates the correction here, noting backbone traffic actually grew at roughly half the “doubling every 100 days” rate. See Fabricated Knowledge.
Tesla’s forward price/earnings was about 215x in mid-May 2026 — GuruFocus recorded 215.24 on 17 May 2026 — against a market capitalisation near $1.6 trillion, a multiple the car business alone does not support. See GuruFocus.
SpaceX's all-stock acquisition of xAI closed in early February 2026, valuing the combined entity at $1.25 trillion (SpaceX about $1 trillion, xAI about $250 billion) — reported as the largest merger on record — and set the stage for SpaceX's June 12, 2026 IPO. Early reporting put the target near $1.5 trillion; by early June 2026 it had been revised up to roughly $1.75 trillion ($135 per share, about a $75 billion raise), which would be the largest public listing in history. See CNBC and Fortune.
NVIDIA’s fiscal-Q1 2027 results, reported 20 May 2026: revenue of $81.6 billion, up 85% year-over-year, with next-quarter guidance of $91.0 billion and gross margin around 75%. See NVIDIA.
Two paired claims. WorldCom/UUNET’s “internet traffic doubling every 100 days” (about 1,000% a year) was repeated widely and even appeared in WorldCom’s 1998 annual report, though analysts including Andrew Odlyzko later showed actual backbone traffic was doubling closer to once a year (Light Reading). OpenAI’s May 2018 “AI and Compute” analysis put the doubling time for compute in the largest training runs at 3.4 months — about 103 days — since 2012 (OpenAI).
Meta fell about 11% after its Q3 2025 report — its worst day since October 2022 — on raised AI capex (CNBC, Oct 2025), then fell roughly 6-7% after its Q1 2026 report on 29 April 2026, when it raised 2026 capex guidance to $125-145 billion (CNBC, Apr 2026).
Epoch AI’s benchmark-controlled analysis finds inference prices for a fixed capability level falling by roughly an order of magnitude per year — about 40x annually for GPT-4-class performance, with the steepest declines after 2024. See Epoch AI.
Bitmain’s Antminer S9 to S17 to S19 to S19 XP line illustrates the sub-annual obsolescence cadence as mining difficulty rose. See D-Central.
Solar-module prices fell from roughly $4 to roughly $1 per watt over 2009-2012 as capacity outran demand. See Our World in Data.
Solyndra’s 2011 collapse — the subject of a Congressional Research Service report — was the best-known of a wave of solar-manufacturer bankruptcies in that period. See Congressional Research Service.
Originally published at Rhyd Media.
Enjoyed this article. Here’s my take. Would love to hear your thoughts.
https://supriyasubramaniam.substack.com/p/the-musk-of-mars?r=3klxxr&utm_medium=ios