For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 780 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
This is a dated book. Every price in it is from late June 2026, and the names that move fastest in this market round-trip thirty percent in a season, so read it as a snapshot of how I'd position, not as a standing recommendation, and certainly not as advice. I hold no edge in markets, and the honest way to write in a domain where you have no edge is to commit your calls in the open, with numbers, where they can be marked against reality later.
One frame runs underneath all of it. Scarcity in a capital buildout comes in two kinds with opposite half-lives: a physics-gated chokepoint that gets harder to clear every generation, and a contract-gated peak that decays the moment new supply or a re-bid arrives. The whole book reduces to three moves. Own the scarcities that deepen. Rent the ones that decay, and only at a price that assumes the peak ends. Refuse to pay a deepening-scarcity multiple for a decaying-scarcity margin, which is the trade the crowd is most eager to sell you near a top. Here are a dozen names against that test.
The one firm that volume-manufactures both leading-edge logic and the advanced packaging that stitches every accelerator together, and the packaging step gets worse with each generation. It runs thirty percent revenue growth at roughly sixty-six percent gross margin, and it trades around twenty-five to twenty-seven times forward earnings, which is cheaper than the chip designers who cannot ship without it. The whole supply chain pays up for the customer and discounts the supplier the customer depends on. At about $432, ten percent off its highs after a near-doubling, I'd buy weakness toward $360-390 rather than chase. Twelve-month base case $480-560; the eighteen-month band is wide ($350-650) because the first real year of capex digestion is the swing. The only cap on conviction is Taiwan, a sovereignty tail rather than a competitive one, which is why this is a core holding and never a maximum bet.
The strange fact about Nvidia in mid-2026 is that the stock got cheaper as the business doubled. Data-center revenue ran about ninety-two percent year-over-year to roughly $75B in a quarter, gross margin held near seventy-five percent, and the forward multiple compressed to the low twenties with a PEG under one-half. The moat that genuinely deepens is the system, not the die: the NVLink and Spectrum-X fabric, the rack-scale integration, the software gravity of CUDA, with networking revenue growing faster than the chips. That layer compounds. The layer that erodes is inference, where a hyperscaler running one model family can tape out a custom part at several times the efficiency, and Broadcom's order book proves they are doing exactly that. So I'd accumulate on dips rather than pound the table: a deepening chokepoint at the frontier, sitting on top of a cyclical peak in inference, at a price that has already stopped pricing perfection. Twelve-month range $210-300, conditioned on data-center growth staying above forty percent and no broad capex cut.
The only company that owns the entire AI stack: its own frontier models, its own competitive silicon, its own hyperscale cloud, its own distribution to billions. The single most under-modeled fact in the stock is that the silicon stopped being internal. Google now rents its accelerators to Anthropic and to Meta and intends to sell them to outside data centers, which turns a cost center into a second compute-revenue line and a cost edge no rival has. Cloud re-accelerated through the year to better than sixty percent growth at roughly a thirty-percent margin, search revenue grew nineteen percent against two years of disruption fear, and the December 2025 antitrust remedy cut the breakup tail without a forced divestiture. After a ninety-percent year it is no longer cheap on price, only cheap on multiple relative to what it is, near the megacap average while compounding faster. I'd add on regulatory or capex-driven pullbacks. Twelve-month range $360-460; the live risks to buy into are the enormous capex intensity and the ad-tech appeal, neither existential.
The credible second pole to Nvidia, and the one selling hyperscalers the custom silicon they most want to own, with the VMware software annuity underneath for ballast. Custom-ASIC revenue cleared $8B in a quarter against a backlog measured in the tens of billions, spanning Google's, Meta's, Microsoft's, and Amazon's in-house parts. At roughly twenty-three times forward it is fairly valued, not cheap, and the revenue is lumpy and concentrated in a handful of programs that get re-bid. I'd buy a capex-fear flush into the high $200s rather than pay $365 for it. It is a cyclical-hearted winner: a real franchise whose quarters will still swing hard with the buildout's mood.
AMD did the hard thing and broke into the accelerator duopoly, with multi-gigawatt commitments anchoring its 2027 parts. The trouble for a buyer is that the contrarian thesis already played out in the tape, up more than a hundred percent on the year, which leaves it priced for flawless execution with no valuation floor to catch a stumble. The cleanest expression here is the pair, long Nvidia against short AMD, to isolate the execution gap rather than bet the beta. Outright, I'd wait for $380-430. Good company, spent setup.
The scarcest physical asset in the buildout is not a chip, it is firm clean power, and Constellation owns the largest fleet of it, with a large block of 2030 output still uncontracted at pre-bidding-war prices. You cannot build sixty gigawatts of nuclear in a hurry at any price, so the repricing optionality on that uncontracted block is close to free, and the hyperscalers signing premium long-tenor deals prove the bid is real. Down roughly a third from its October 2025 peak to about $270 and around twenty-three times forward, it is the rare power name where the moat is genuinely irreplaceable on any near horizon. I'd start half size and let a signed contract, not a chart, add. Best entry $230-245; twelve-month range $280-360, with the excess return front-loaded into the next few years of repricing.
The single best raw valuation on the board and the only full-stack AI platform in China, silicon to model to cloud, at roughly fourteen times earnings with cloud growing forty percent, which effectively hands you the commerce business for free. The catch is the freshest catalyst, and it points down: an allegation in late June around the Qwen models and a sanctions tail that threatens the very moat the bull case rests on, against a position the whole street already owns. So this is a sub-one-percent starter near multi-year lows, with room to add on a flush, not a back-up-the-truck conviction long. I'd hold the Hong Kong listing rather than the American shares to sidestep the delisting wrinkle. Twelve-month range $90-160, genuinely binary on the August print and on whether the sanctions noise escalates.
A contracted backlog near $100B, almost twice the market cap, with all four frontier labs signed. It is also a leveraged landlord of depreciating hardware on a financing treadmill, where the chips obsolesce faster than they amortize and a single renegotiated mega-contract can snap the income statement before it turns. That makes it a solvency instrument wearing a growth-stock costume. Size it like a binary option, small, and enter only after a clean quarter and a refinancing clear into the $64-75 zone. The upside re-rate and the impairment are both live; do not pretend otherwise by sizing it as a normal long.
Now public after the largest listing in history, trading around $153 for a roughly two-trillion-dollar cap, which is north of a hundred times sales on a business still running deep losses, much of the cash burned on the embedded AI venture stapled to it at a related-party mark. The real asset is the space and connectivity monopoly; you are overpaying for the venture premium bolted to it. The supply picture is the tell: the lockup is front-loaded, with a large tranche releasing at the coming quarterly print and steady tranches through the autumn. I'd let that supply mark it down and buy the monopoly into a $100-120 flush, small, rather than chase the scarcity pop at $153.
Around ninety percent of a one-and-a-half-trillion-dollar valuation is optionality on robotaxis and humanoid robots that carry almost no revenue before 2027, sitting on a measurable car business whose earnings are falling. A trailing multiple in the hundreds on a shrinking base is not a price I can defend long, and the same optionality makes it a poor short, since narrative can squeeze it for quarters. No position. If the autonomy thesis ever converts to recognized revenue, that is the moment to re-underwrite it, not now.
The headline avoid, and the cleanest example of the frame. Micron posts gross margins near eighty-five percent on memory, software-grade numbers on a physical commodity, because supply is briefly tight, and the stock has run nine to eleven times off its low to all-time highs around $1,130. The low forward multiple reads cheap and is the opposite: it is the market's discount mechanism marking earnings it knows are about to roll, and to buy you must believe both that the peak persists and that the multiple re-rates up, which is being paid twice on one bet. The June drawdown despite a sold-out 2026 is smart money pricing the 2027 turn. No edge long. If you want the cycle, rent it through a chip index, not a single-name peak.
A genuine non-GPU architecture and a real model-serving niche, freshly public after an IPO that opened near $350 and round-tripped to about $182 inside six weeks. The arithmetic is the problem: roughly forty-six times forward sales, negative operating margins, and the great majority of revenue from a single related party in the Gulf. That is a venture lottery ticket trading on a public exchange. With the lockup overhang largely spent it is not a short either; it is a pass, or a tiny starter only after a clean post-lockup quarter that quantifies the marquee contract.
If I could only hold three, they would be TSMC, Alphabet, and Constellation — one chokepoint in silicon, one in integration, one in power, none of them the consensus momentum trade and each cheaper than what it actually is.
TSMC, because it captures rising value no matter which model or which chip wins, at a multiple below the firms it supplies, with a tail that is geopolitical rather than competitive. Alphabet, because the market still prices it as a search company with a cloud also-ran when it is the only full-stack AI company there is, now renting its own silicon to its rivals, near an average multiple while it compounds above average. Constellation, because firm clean power is the bottleneck that the chip race cannot route around, and it owns the irreplaceable asset with the repricing still ahead of it.
Nvidia and Alibaba are the next two off the bench, the first a deepening chokepoint I'd rather buy on a dip than at the high, the second the best price on the board if you can stomach a live sanctions tail. The memory names are the ones I'd be selling into strength.
The discipline that generated every line above is the same: ask whether the scarce thing gets scarcer as demand grows, or whether demand invites the supply that competes it away. Own the first kind through the cycle. Rent the second, cheaply, and never confuse the two. The market keeps stamping the whole chain with one word. The work is reading the half-life it forgot to print.
Not investment advice. Prices as of 27 June 2026; figures from public filings and market data as of that date. I have no position-management mandate and no edge in markets — these are reasoned scenarios committed in the open so they can be checked.