AI made building a valuation model cheap, and made checking it the job. This is how I work with Claude, from framing what a price already assumes to shipping a live model a reader can challenge, with a human call at every step. In a model, the second-worst outcome is a wrong number. The worst is a right number removed on purpose.

An account of building an institutional-grade valuation model with AI, and of the eight times it was wrong. Not one was an arithmetic error. Every one was a claim about the output — a stale figure from an earlier build, a crossed reference, a share basis that didn't match the line beside it. What caught them was a regression harness, a second AI system with no stake in the first one's conclusions, and review. The work didn't shrink; it moved from construction to verification.

The principles that guide my work as an investor and builder. How capital allocation determines whether value gets created or destroyed. Price vs value, leverage, risk, verification, and why the discipline of the work matters more than getting lucky.

ETH on exchanges has fallen to a near-decade low, but the coins moved into staking (~42M since 2020) rather than out of the market. This follow-up to Parts 1 and 2 finds the exchange balance has tracked price rather than led it, explaining about 2% of the next 90 days’ return.

IREN’s October share sales, about $630m at roughly $36 a share, lifted book value from $12.57 to $13.54 a share. This one-page follow-up to my Texas pause analysis explains why selling above book doesn’t mean selling above value, and why losses, net of stock compensation, pull book back down by about $0.37 a share each quarter.

A per-share look at what a prolonged Texas grid pause would cost IREN: roughly $2–8 a share, partly offset by pacing construction. How the remaining ~$8bn is funded, debt or equity, matters as much as the pause itself.

Since August, Texas has paused the grid study and switch-on of new large data centers, along with their environmental permits, while the state audits a queue of roughly 474 GW. These two slides map how a data center of 75 MW and up gets onto the ERCOT grid, where each step stands today, and who can change it: the pauses rest on the governor’s directives, while ending the data center tax breaks would take a new law. It’s a general guide to the process rather than a view on any company, built from ERCOT, state agency and statutory sources as of November 2026.

A test of whether the market is pricing the neoclouds’ business model or each company’s balance sheet. Net of the Nasdaq, roughly 65–80% of CRWV, NBIS and IREN’s daily moves are shared, and SPCX, the better-capitalized rival, barely moves with them. The one early sign of differentiation is NBIS: on days CoreWeave’s 2032 bonds sold off since October, it fell ~1% against ~3% for CRWV, though that rests on only 5–9 sessions and isn’t a bottom call.

Ethereum keeps widening its “road” before traffic arrives, so the toll that gets burned has fallen ~99% since 2024 and now offsets only ~2% of new ETH issuance. Using on-chain data from the last three gas-limit increases, I model what Glamsterdam’s added capacity does to fees: unless traffic roughly doubles first, the toll goes to near zero. Fees won’t carry ETH’s price by design, so the value has to rest on staking and ETH as money, which the next installment tests.

Part 2 tests whether usage is turning into demand to own ETH, through the bond (staking) and the money (ETH held as a reserve). Value secured on Ethereum and long-term holding have both grown since 2024, yet the price hasn’t followed, and the deck sets out what that means for today’s valuation.

Heading into IREN’s November 5 print, this update re-solves my valuation model for the cost of capital at the ~$24 close. The price implies a WACC of ~12.3%, about 1.3 points above the ~11% modeled, a premium that widened even as Treasuries eased. Taking the options market’s implied move for earnings week, a 1σ move to ~$19 or ~$30 is worth about half a point of WACC either way, while a two-quarter Horizon delay costs only ~$1.4 a share.

A read on Intel after its Q3 2026 report. At ~$66, the stock needs ~$2.20–3.30 of annual EPS at 20–30x, against a ~$2.08 Q3 run rate and ~$1.00 implied by the Q4 guide. A revenue bridge built from the 8-K and the earnings call shows the guide-down is a client destock, larger than the company's entire ~$1bn decline, while data center keeps growing, so there's no urgency to own the shares before the January guide, in our view.

Meta now spends ~89% of its operating cash flow on capex and has ~$810bn committed beyond its balance sheet, as financing shifts from cash to debt to new equity: ~$37.5bn priced on October 26, with a ~$60bn ATM to follow. A reverse DCF shows that at ~$631 the price asks for ~17% revenue growth a year for a decade, against ~12% at the 2021 peak, and the higher hurdle rate explains all of the gap. Built entirely from SEC filings and audited against every earnings release since 2019, with the 2022 metaverse drawdown as the reference case.

As credit markets repriced neocloud debt, I asked what that means for IREN's valuation, its funding plan and its downside. The analysis ties FINRA TRACE bond prices and Treasury yields to the WACC implied by IREN's share price, reconciles management's funding guidance with my own model, and stress-tests how much the valuation can absorb. Every figure is a full recalculation, audited and sourced to primary filings.

A survey of the cost of capital at CoreWeave, IREN, Nebius and SpaceX, prompted by CoreWeave's ~15% bond yields and SpaceX's third-quarter results. Against an investment-grade contract, the neoclouds borrow about as cheaply as SpaceX; the gap opens on the dollar no contract backs, which costs them ~13–22% against ~10–14% for SpaceX. The market is pricing that marginal dollar rather than the average cost of capital, and SpaceX's edge lies in how much it has already raised, not in what it pays.

As credit markets repriced neocloud debt, I asked what that means for IREN's valuation and its funding plan. The analysis ties FINRA TRACE bond prices and Treasury yields to the WACC implied by IREN's share price, then reconciles management's FY27 funding guidance with my own model. Every figure is a full recalculation, audited and sourced to primary filings.

An October update to my IREN analysis after the stock fell ~16% in a week to ~$35. Re-running the same model shows what the new price discounts, about 11% of the recent pricing gain reversing or GPUs lasting ~4.5 years instead of 5, and how quickly value falls beyond that: roughly $8 a share per 10 points of pricing reversal. Two sensitivity matrices map the downside across pricing, GPU life, cost of capital and utilization.

An independent valuation of IREN testing whether scarce powered capacity in Texas helps or hurts the company. At ~$42 a share in October 2026, the price assumes today's pricing, utilization and financing persist, and the bigger risks are contract pricing, GPU lives and funding a ~$55bn build rather than the Texas queue.

A two-slide analysis of what ETH's price actually rests on. Fees support roughly 9% of today's price. Read through a security-budget rule, the price instead needs Ethereum to secure ~$2.3T of stablecoins and tokenized assets by 2035. A second lens converts a ~$4T agentic revenue estimate into the balances behind it, and finds today's price needs only ~0.5% of them on Ethereum, if the assets follow the revenue.

Sharplink holds almost nothing but ETH, yet its stock has traded below the value of those net assets on every session for a year. Rebuilt from primary filings, the analysis shows the discount's swings fade within weeks, but only back to about 0.8× NAV, while ETH drives about three-quarters of the stock's moves. With the market not closing the gap, the question shifts to management and what it can do to grow ETH per share.

A daily log-return decomposition of the LINEA token, splitting every price move into four terms that add exactly: ETH, network fees, the valuation multiple, and circulating supply. Built from primary on-chain data — 252,887 transfer events replayed from the genesis mint, supply reconciled to the token contract to the wei, and network fees computed block by block — with no explorer labels or aggregator figures. The result: since the dual burn went live, 38 million tokens burned against 16 billion unlocked, and a market paying 176× fees for adoption that has not yet arrived.

Sharplink holds ETH, but owning the stock is not the same as owning the asset. Decomposing thirteen months of returns separates what came from ETH, what came from the multiple the market puts on net asset value, and what leaked away in dilution. The mix depends on how long you hold — the multiple drove half of a single day's move and a third of the period, because it mean-reverts and ETH doesn't.

IREN is two investments stacked: a data centre building, and a decision to fill it with hardware rather than lease it out. With a third of its capital in real estate and more than 80% of announced capacity still to be built, the market is paying for contract pricing less than where the company signs today.

Three interactive teaching tools I built with Claude for two math classes I guest taught in the local public school system. Students drag points to see average rate of change, slide parabolas to see how equations reshape graphs, and watch a looping walkthrough of polynomial skills.

Digital asset treasury companies get valued on ETH per share, a metric that rises on any buyback at any price and never registers a dollar of operating expense. This model puts NAV per share beside it and prices the warrant stack properly — treasury stock method where warrants are in the money, a Black-Scholes deduction where they are not. The anchor is what SBET is worth if management never touches the capital markets again; everything above that line is what you are paying for the machine.

Best viewed on desktop

A sum-of-the-parts DCF that values ETH as programmable infrastructure, modeling composability, restaking, and agentic settlement as calibrated overlap rather than a comp-multiple guess. Built as a single dial-able valuation with no bull/base/bear scenarios. Simply adjust any assumption and watch one number, value per ETH, update live.

Best viewed on desktop

SpaceX defies a single valuation line. This sum-of-the-parts framework breaks the business into its constituent segments and lets you stress-test the assumptions that drive the range. Rebuilt from Excel into an interactive app you can actually challenge.

Best viewed on desktop

A DCF and fee-multiple framework for the Linea protocol. This is the kind of model that usually lives buried in a spreadsheet, rebuilt as an interactive tool you can actually steer.

Best viewed on desktop

A browser-based replica of the HP 12c RPN financial calculator, the desk classic a generation of analysts learned on, including yours truly, faithfully rebuilt for the web.

Best viewed on desktop

A low sales multiple can still be expensive. This interactive lesson shows why price-to-sales and annualized revenue mislead for capital-intensive companies: three hypothetical companies with the same sales, growth and margin turn out to be worth roughly 4x apart, because the multiple hides how much capital growth requires and what it earns. Step through the frames, then set a market price and see what return on new capital it assumes.

A company’s average return on capital blends old and new investment, so when spending surges, the average lags what the new dollars actually earn. This lesson applies the idea to Microsoft and Oracle using their 10-K and 10-Q filings, and includes a back-solve tool for what new capital must earn to hold today’s ROIC.

Losses and gains aren’t symmetric: a 50% drawdown needs a 100% gain just to get back to even, and at a 6–10% annual return, that recovery could take roughly 7–12 years. The required gain follows L / (1 − L), so the climb back steepens quickly as losses deepen, which is why avoiding the large drawdown often matters more than capturing the next rally. I put this together after the recent declines in the neocloud names, but the math applies to any position or portfolio.

A prepayment is a loan from the customer, repaid in service: in an illustrative $100m GPU purchase that is 50% prepaid, the company funds ~$50m at the peak instead of ~$100m, but the capex is still $100m. IREN's recent contracts have ~45-55% of GPU capex payable upfront (FY26 results), which eases the build-phase funding gap without making the build free.

Whether book life reflects the economic life of the hardware is now central to the AI infrastructure trade. Nowhere more so than in neoclouds, the specialized cloud providers that help make AI possible.

Five companies — Google, Microsoft, Meta, Amazon, Oracle — are pouring unprecedented capital into AI infrastructure. Growth isn't the question. Value creation is.

Valuing Ethereum's potential in the agentic economy. Our approach is a five segment valuation model that treats ETH as programmable infrastructure. Includes sensitivities on what AI agent settlement could do to the valuation.

SpaceX is more than one business, and each carries its own risk. This presentation introduces a sum-of-the-parts valuation of Space, Connectivity, AI, and Expansion, each priced at its own cost of capital rather than blended into a single company-wide discount rate.