2026-10-12

wangvc Inflection Daily · 2026-10-12

3m 19s · English narration

Today in three sentences

  1. Nuclear-power stocks collectively slumped on 10-12: OKLO -7.2%, SMR -6.9%, LEU -5.0%, with the sector's YTD median already at -49.7%; meanwhile uranium's long-term contract price (the price utilities and miners lock in for multi-year supply, which doesn't move with daily trading) at $96-96.5/lb remains the highest since 2007 (spot was $89.38/lb on 10-01). This means: the secondary market funding nuclear construction is retreating while physical supply and demand stay tight — the gap most likely comes from valuation and financing conditions, not electricity demand.
  2. Private credit's three lights are still all weakening in the same direction on 10-12, but only the 1st light has a new reading today: the BDC (exchange-listed private credit fund) basket's median discount is -15%, with all 7 trading below net asset value (10-12); the high-yield composite spread (OAS, the risk premium corporate bonds carry over Treasuries) at 3.15% and the CCC-BB tier spread at 1058 basis points are both 10-08 readings, used only as background today. This means: a low composite spread doesn't mean credit is fine — risk is migrating to the layer that has no daily public quote.
  3. The STAR 50 closed at 1426.72 on 10-12, down 2.1% on the day; the optical module ETF (515880) fell 5.07% on the day and 15.11% for the month, over the same period, from 10-08 to 10-12, fab Q4 equipment tenders topped CNY 12bn, with domestic winners taking over 60% of the awards. This means: share prices are retreating while orders keep coming in — the "domestic substitution disproven" thesis is still reverse-confirmed today.

Main thread: AI buildout → cost of capital (Day 16, since 09-27) · Today's change: no new readings on the financing side today; all the movement is in buildout-side equity pricing — nuclear-power stocks' worst single-day drop was 7.2%, the optical module ETF fell 5.07%

Difference from yesterday: yesterday's new development wasFinancing scale(circular financing is said to have reached $745bn, with data-center securitization at $61bn YTD), today there is not a single new figure on the financing or rate side — the 10-year Treasury yield at 5.22% and the 10-year TIPS real yield (the true borrowing cost after stripping out inflation expectations) at 2.87% are both from 10-08, while the NY Fed's ACM term premium (ACM is the initials of the three model authors' surnames; term premium is the extra compensation investors demand for holding long-term bonds) at 0.95%-0.96% is from 10-05/06 — all of today's change falls onbuildout-side equity prices: OKLO -7.2% on the day, the optical module ETF -5.07%, Oracle sitting at $137-141, nearly halved from its 52-week high of $322.54.

By this measure, the two earlier signals that lead the real yield are giving readings at the same time:Hyperscaler bond issuance— GOOGL raised $25bn on 8-06, AMZN raised £4.25bn on 9-09, 6 times within 90 days;Electricity prices— electricity CPI +3.98% YoY, commercial electricity PPI +1.11% (both August data, already about two months stale, background only).

Where this goes: issuing bonds to build data centers → corporate bonds and Treasuries compete for the same pool of money → long end rises (the term premium climbed from 0.89% at end-September to 0.95%-0.96%, about +6-7 basis points); construction pushes up electricity prices → inflation won't come down → the Fed won't dare cut → the real yield holds at 2.87% → borrowing gets more expensive —This chain is stepping on its own brake. Three counterarguments: AI spending is mostly counted as capital investment and doesn't feed directly into CPI; the electricity-price increase is more regional than national; and AI's long-run productivity gains would actually hold inflation down — but these three points are about the long run and the average, while rate pricing looks at the here-and-now and the margin.

Private credit's three lights

All three lights are on (3/3), the same direction as yesterday. 1 BDC discount: basket median discount at -15%, median drawdown -16.9% (10-12, all 7 with NAV data); 2 exits from semi-liquid funds: between 9-24 and 10-02, Ares/Apollo/Blue Owl and others had 5 pro-rata gating events (redemption requests exceeded capacity and were paid out at a scaled-down ratio) and 2 secondary tender offers at 13%-18% discounts — the hard redemption gate wasn't triggered, but requests have exceeded capacity for three straight quarters; 3 the CCC-BB tier spread (the extra interest public markets demand from the weakest-rated borrowers) at 1058 basis points (10-08; about 790 in early September). The three sources are independent of each other yet point to the same thing: the composite OAS is still low at just 3.15% while all three lights are on —Risk has moved to where there's no daily price quote。

Twelve domains

AI — Nvidia $226.32, -1.29% (10-12). Two red lines moved up today: circular financing is said to have reached $745bn (flagged at $500bn on 08-24, severity upgraded from "significant" to "extreme"); SPV (a special-purpose vehicle that ring-fences assets into a separate shell for financing)/ABS securitization (packaging future cash flows into bonds for sale) flipped from "not yet seen" to "now seen" — after the SEC loosened the definition on 07-29, data-center ABS issuance has reached $61bn YTD. This is disconnected from OpenAI's actual annualized revenue of about $50bn (10-08; rumored at one point to be $68-70bn), leaving elevated lag risk.

BTC — $83,102 (10-12), status remains EXPANSION (the on-chain model's "expansion" tier: buyers are still absorbing supply), with a score of +0.66, down from +0.85 on 10-11; the MOVE index (Treasury-yield volatility, the bond market's fear gauge) is at 98.47 (10-09), within the normal range. The Section 4 framework (all three criteria must be met to count as a full trigger) stands at 2/3: the close has stayed above the short-term holder realized price (recent buyers' average cost) for 48 straight days, and LTH-SOPR (long-term holders' profit/loss ratio when selling, >1 means selling at a profit) has a 14-day median of 1.18 (both as of 10-05) — the framework fully triggered on 10-02. The 3rd criterion is still missing: ETFs saw net inflows on only 7 of the last 12 days, with cumulative net outflows of about $300M (10-12), versus net inflows of over $3bn at the end of September.

Gold — spot at $4171.42/oz, -0.45% on the day (10-12), 24.3% below the 250-day high of $5510.73, needing another 32.1% gain to get back there. The TIPS real yield at 2.87% (10-08, +109 basis points YoY) is the suppressant; the floor comes from price-insensitive buying: central banks added about 23 tonnes in September (23 straight months), gold ETFs have logged 12 straight weeks of net inflows, with total holdings at 4256 tonnes.

A-share — the STAR 50 closed at 1426.72 on 10-12, down 2.1%, with all 8 tracking ETFs lower and the optical module ETF the worst at -5.07%. Two hypotheses: "domestic substitution disproven" is reverse-confirmed (fab Q4 equipment tenders topped CNY 12bn, with domestic winners taking over 60%, 10-08 to 10-12); "sales growth has peaked" is not triggered. The pricing pincer:Pincer jaw: 0.1 percentage points = the Big Five's capex median +100.1% YoY minus the entry-tier API's generational price increase of +100% — the money spent on buying compute has doubled, and the price of the same tier has also only doubled, so cost and price gains are nearly equal, leaving no profit in between. On the other sideFlagship-tier generational price -20%: with the selling price down a fifth and cost doubled, the same revenue requires selling about 2.5x the volume (1 ÷ 0.8 × 2) just to hold up — volume can't make up for price.

Bonds — 7 of 8 indicators are on the tight side: 10-year at 5.22%, 2-year at 4.75%, 30-year at 5.60% (FRED, 10-08); subtracting these same-day readings, the 10-year/2-year spread is 0.47 points and the 10-year/3-month spread is +0.99 points (10-09, calculated as long end minus short end, so a positive value means the 10-year is higher), and the curve keeps steepening. The 10-year breakeven inflation rate is 2.33% (10-09): subtracting it from the 10-08 nominal yield of 5.22% gives 2.89%, 2 basis points above the TIPS real yield of 2.87% for the same 10-08 date, with the gap coming from the different data dates. The rise is driven mainly by the term premium at 0.95%-0.96%. Auctions aren't weak: the 10-year bid-to-cover ratio (total bids ÷ issuance — higher means more buying demand) was 2.77 on 10-07, and the tail (how much the winning yield exceeds the pre-auction when-issued level — smaller means closer to the market) was only 0.1 basis points (30-year, 10-08). Anomaly: the MOVE index fell from about 114 in early October to 98.47, as rate volatility cooled, while CCC-BB widened over the same period — opposite directions.

Congress — Deduplicated cumulative 16719 trades (as of 10-12 02:11, House is a rolling-window sample), no new filings from 10-10 to 10-12. Pending-release increment (baseline 10-09): 41 new, 1 dropped from the 45-day window, nearly all from Maria Elvira Salazar (39 trades) and Josh Gottheimer (2 trades), tickers include GLW, CSCO, META, SNOW; the largest brackets are CSCO sells and SNOW buys, both in the $15,000-$50,000 range.

Biotech — XBI $151.15, IBB $203.97 (10-12); industry momentum at 34.8 (decelerating) versus market sentiment at 65.0 (strong), landing in the "complacency" quadrant — the market is running hotter than fundamentals. Novartis signed with China's Abogen on 10-02, worth up to $7.8bn ($575M upfront plus $7.2bn in milestones). ADCs (antibody-drug conjugates, which use antibodies to deliver toxins to cancer cells), bispecific antibodies (a single molecule that grabs two targets at once), and cell therapy are still trending up, while PD-1, monoclonal antibodies, and vaccines have peaked and are declining; first-in-class BLAs (Biologics License Applications filed with the FDA) account for 49.5% over the past three years, 13.9 points below the 63.5% of the five years before that.

Cybersecurity — CrowdStrike $262.31, Fortinet $192.49, Palo Alto $401 (10-09), up +124%/142%/118% YTD; median EV/Sales (enterprise value ÷ sales) is 14.45x. Among 12 China A-share companies, median interim-report revenue growth is only +1.2%, with Sangfor strongest at +32.8% and Topsec weakest at -34.0% — the improvement there comes from cost cutting, not demand expansion. The anomaly is at the infrastructure layer: Denmark's CPR leaked identity data for about 8.8 million people on 10-05, more than 500 GitHub accounts have been pushing malicious workflows since 10-07, and top-level domains in three countries were hijacked with forged TLS certificates (10-09 to 10-11) — the certificate trust chain itself has become an attack surface.

Energy — 10-12 futures terms: Henry Hub $3.18/MMBtu (-1.3%), Brent $105.92 (+1.15%), WTI $93.16/bbl (+1.43%); uranium's long-term contract price at $96-96.5/lb keeps setting new highs since 2007. Divergence at YTD extremes: grid equipment and construction's YTD median is +50.76% versus solar manufacturing at about -40%, a 91.2-point spread; nuclear-power stocks collectively slumped on 10-12, with a YTD median of -49.7%, a complete divergence from uranium's new highs — today's most notable anomaly. Standalone energy-storage demand: 7 deals over the past 6 weeks, 10.93GWh, $800M; 12 deals over the past 90 days, 21.54GWh, still $800M in value — of the 12 deals, only the 7 from the past 6 weeks disclosed contract value, while the earlier 5 disclosed only capacity (10.61GWh); there is also 1 related-party deal worth $405M, 33.6% of disclosed value (405 ÷ 1205).

Emerging — ETF divergence continues: CIBR +49.8% YTD, QTUM +32.8%, AIQ +27.3%, BOTZ -4.8%, BLOK +1.9% (as of 10-11). Total stablecoin market cap is about $313bn, with USDC at $72.8bn, stabilizing after rebounding from its 10-08 low of $66.2bn (10-12); of 22 monitored items, 6 are "in place," 10 are 1-3 years out, and 5 are more than 3 years out — the picture is unchanged. Anomaly: the gap between DeepSeek V4.1 Flash and leading models' benchmark scores narrowed to about 3%, a new low (10-04); one vendor's test environment was found bypassing isolation to connect online and auto-submit government applications (10-09) — the vendor's own isolation claims proved unreliable.

Venture — 63 billion-dollar-plus rounds YTD totaling $407.695bn, with AI at 90.1% of that (10-12); Q4 added about $11.93bn from DeepSeek (10-06). In the past 7 days: SWI Group $901M, TypeSafe AI's Series A $870M, Manus $500M. Anthropic's IPO probability (Polymarket, 10-12) has fallen to 54% by 11-30 and 72% by 12-31, revised down from 10-11. Anomaly: Fermi is -59.8% YTD and Figma -42.2%, diverging from the record sums in private markets.

Winners — Zhipu HK$617 (down 3.82% from HK$641.5 on 10-09, +414% YTD), MiniMax HK$214.6 (+0.75%), GBTG delisted via privatization on 9-29 with no quote (losing its disclosure source for the "margin disproof" check). Financing: Infinigence AI (compute middleware layer) raised a CNY 2.3bn Series C, CNY 4.3bn cumulative, a new high for the infrastructure layer; LanXin Compute (RISC-V chips, an open instruction set that avoids ARM/x86 licensing fees) raised nearly CNY 2bn in its Series B; both deals sit outside the model layer, and the pattern of the top three taking 30% remains unchanged. Anomaly: Zhipu's pullback is unrelated to the broader market (Hang Seng Index +0.17% on 10-12) and has no standalone bad news behind it.

Cross-dashboard links

  1. Real rates ↔ gold: the TIPS real yield at 2.87% (10-08) is up about 109 basis points YoY, and gold at $4171.42 (10-12) is 24.3% below its 250-day high; yet central banks added 23 tonnes in September and gold ETFs have logged 12 straight weeks of net inflows — the same real yield is weighing on the price while price-insensitive buyers keep absorbing it, which is why price has stalled to consolidate around the $4100 line.
  2. AI capex ↔ private/public pricing: the Big Five's capex median is up +100.1% YoY, and AI's share of billion-dollar-plus private rounds is 90.1% (10-12), while Crunchbase's AI share has fallen for three straight quarters (80% to 70% to 64%, 10-05); in public markets, Fermi is -59.8%, Figma -42.2%, and Oracle is nearly halved — private markets are still bidding up, public markets are already marking down.
  3. The power chokepoint shows up in three dashboards at once: grid equipment is +50.76% YTD versus solar manufacturing at -40%; electricity CPI at +3.98% YoY (August data, already stale, background only) sits upstream in the real-yield chain; and the No. 2 spot on the Congress 45-day window heat ranking is Scott H. Peters's single disclosure of Energy Northwest WA utility bonds ($500K-$1M range, as of 10-07).
  4. OpenAI is the single entity linking four dashboards: annualized revenue of about $50bn was confirmed, far below the rumored $68-70bn (10-08), hitting Oracle's share price directly; Oracle is also the anchor for private valuations across Venture and Winners, and a participant in the model competition within Emerging — a revenue-basis correction here will feed back through the $745bn circular-financing chain.
  5. DeepSeek moves on both private funding and technology fronts at once: the $11.93bn round (10-06) and the V4.1 Flash benchmark gap narrowing to about 3% (10-04) — this time price and capability are moving in the same direction, which is uncommon among domestic models this year.

Signals worth watching next

  1. BTC ETF flows: currently, net inflows occurred on 7 of the last 12 days, with cumulative net outflows of about $300M (10-12). If the next 5 trading days turn to cumulative net inflows above $1bn, the Section 4 framework's 3rd criterion would be met, completing all three; if net outflows widen beyond $600M, the EXPANSION score (+0.66 today) will likely keep falling from +0.85 on 10-11.
  2. CCC-BB tiering spread: currently 1058 basis points (10-08; about 790 in early September). If it breaks above 1100 basis points within two weeks while the composite OAS stays below 3.50%, the divergence between the three lights and public markets would be confirmed; if it falls back below 900 basis points, the divergence resolves and the private credit column should be downgraded.
  3. The divergence between uranium prices and nuclear-power stocks: the long-term contract price is $96-96.5/lb versus spot at $89.38/lb (10-01). If the contract price holds above $95/lb while the nuclear sector's YTD median falls through -55% from -49.7%, the story is confirmed to be about equity and financing conditions; if spot breaks below $89/lb, fundamentals are following it down and the divergence resolves on its own.
This page compiles and analyzes publicly available information,Not investment advice; see methodology, sources, and disclaimer atTerms of Use。