TSMC checking power availability and customer finances before adding capacity is the perfect lens on your CoreWeave point. Goldman, as cited in 404K Semi-AI's Oct 20 brief, projects a 15.3% unlevered after-tax IRR for CoreWeave even while raising its target to $106. With some CoreWeave bonds yielding above 16%, new debt costs more than the assets are expected to earn. TSMC's caution looks wiser every week.
A timeline on the CapEx figure: the 2026 range is the budget CFO Wendell Huang gave when he raised it on the July 16 call. TSMC's Q3 management report shows actual spending of $11.1B in Q1, $15.7B in Q2 and $13.8B in Q3, so $40.6B through September, with Q3 a bit below Q2. I didn't find the full-year range restated in the Q3 slides or the management report, so whether 2026 lands inside it gets settled when the Q4 actuals come out.
The Q4 guide is what stopped me — $52.3–$54.3B implies about 58% growth at the midpoint, right after a 41% quarter, and HPC has gone from 42% to 68% of revenue in three years. That mix shift makes CoreWeave's 16% bond yields feel like someone else's problem for TSMC, at least for now. Do you think that changes if overseas capacity keeps diluting gross margin by the 3–4 points they flagged?
Good framing, and the numbers make the contrast do the work: TSMC sits on ~$126B of cash against ~$32B of interest-bearing debt with a 59% operating margin, so it is on the lending side of the AI debt problem, not the borrowing side. The skeptical read is that the debt problem can still reach TSMC second-hand. The $100-104B CapEx for 2026 is only as safe as the balance sheets of the customers behind it, and your own reporting notes TSMC checks customers' financials before allocating capacity, which tells you where management thinks the risk sits. Neocloud leverage is the canary, but the coal mine is TSMC's order book. Worth watching whether the AI-accelerator mid-to-high-50% CAGR guidance through 2029 survives a CoreWeave-style repricing, because that number is doing a lot of work in the bull case.
TSMC verifying customers' balance sheets and electricity supply before committing capacity is exactly the diligence retail investors skip.
From my seat with salaried Indian professionals, I keep seeing the same pattern: chasing the AI theme through thematic funds and PMS at peak narrative, without realising they are underwriting a capex cycle financed by debt now yielding 16%.
The chipmaker runs the credit check. The investor buying the story does not.
The CoreWeave contrast is the most useful part of this. Same AI demand, completely different risk: TSMC is funding $60 to 64B of capex out of operating cash flow, with about $126B of cash against roughly $32B of debt, while some of the borrowers building the same capacity are paying 16% plus.
I also liked the detail that TSMC checks customers' power availability and financial strength before adding capacity. That's credit underwriting done by a supplier, and it's probably one of the better early-warning signals for when this cycle turns.
The part about TSMC checking customer balance sheets and power availability before adding capacity stands out, because it puts a buyer with roughly $126 billion in cash ahead of the credit stress. If CoreWeave bonds yielding over 16% mean the weakest buyers drop out of the queue, TSMC's order book could keep looking strong while the neocloud tier shrinks. The headline numbers might hide that shift for several quarters.
TSMC checking power availability and customer finances before adding capacity is the perfect lens on your CoreWeave point. Goldman, as cited in 404K Semi-AI's Oct 20 brief, projects a 15.3% unlevered after-tax IRR for CoreWeave even while raising its target to $106. With some CoreWeave bonds yielding above 16%, new debt costs more than the assets are expected to earn. TSMC's caution looks wiser every week.
A timeline on the CapEx figure: the 2026 range is the budget CFO Wendell Huang gave when he raised it on the July 16 call. TSMC's Q3 management report shows actual spending of $11.1B in Q1, $15.7B in Q2 and $13.8B in Q3, so $40.6B through September, with Q3 a bit below Q2. I didn't find the full-year range restated in the Q3 slides or the management report, so whether 2026 lands inside it gets settled when the Q4 actuals come out.
The Q4 guide is what stopped me — $52.3–$54.3B implies about 58% growth at the midpoint, right after a 41% quarter, and HPC has gone from 42% to 68% of revenue in three years. That mix shift makes CoreWeave's 16% bond yields feel like someone else's problem for TSMC, at least for now. Do you think that changes if overseas capacity keeps diluting gross margin by the 3–4 points they flagged?
Good framing, and the numbers make the contrast do the work: TSMC sits on ~$126B of cash against ~$32B of interest-bearing debt with a 59% operating margin, so it is on the lending side of the AI debt problem, not the borrowing side. The skeptical read is that the debt problem can still reach TSMC second-hand. The $100-104B CapEx for 2026 is only as safe as the balance sheets of the customers behind it, and your own reporting notes TSMC checks customers' financials before allocating capacity, which tells you where management thinks the risk sits. Neocloud leverage is the canary, but the coal mine is TSMC's order book. Worth watching whether the AI-accelerator mid-to-high-50% CAGR guidance through 2029 survives a CoreWeave-style repricing, because that number is doing a lot of work in the bull case.
TSMC verifying customers' balance sheets and electricity supply before committing capacity is exactly the diligence retail investors skip.
From my seat with salaried Indian professionals, I keep seeing the same pattern: chasing the AI theme through thematic funds and PMS at peak narrative, without realising they are underwriting a capex cycle financed by debt now yielding 16%.
The chipmaker runs the credit check. The investor buying the story does not.
The CoreWeave contrast is the most useful part of this. Same AI demand, completely different risk: TSMC is funding $60 to 64B of capex out of operating cash flow, with about $126B of cash against roughly $32B of debt, while some of the borrowers building the same capacity are paying 16% plus.
I also liked the detail that TSMC checks customers' power availability and financial strength before adding capacity. That's credit underwriting done by a supplier, and it's probably one of the better early-warning signals for when this cycle turns.
The part about TSMC checking customer balance sheets and power availability before adding capacity stands out, because it puts a buyer with roughly $126 billion in cash ahead of the credit stress. If CoreWeave bonds yielding over 16% mean the weakest buyers drop out of the queue, TSMC's order book could keep looking strong while the neocloud tier shrinks. The headline numbers might hide that shift for several quarters.