Commerce cleared Chinese ghost firms to handle Nvidia's hottest chips — BagBot
Weekly Recap
Commerce cleared Chinese ghost firms to handle Nvidia's hottest chips
Strong jobs data revived Fed hike fears and sent tech down 3.4% for the week, but Friday's 4% chip bounce softened the damage after Commerce's H200 export approvals raised sanctions questions.
TL;DR
Commerce approved Chinese shell companies to handle Nvidia H200 chips while 600,000 sit in Hong Kong — Double or Nothing's NVDA put printed +$9,500
Jobs data sent Fed hike odds to 78%, yields to 4.86%, and crushed small caps for -$9,629 across IWM trades as rate fears repriced the whole week
AI giants are carrying $4.5 trillion in off-book data center commitments — Anthropic's unreported obligations are 100x its balance sheet liabilities
Chips bounced 4% Friday after Thursday's 12% rout but tech still closed down 3.4% for the week; platform lost -$4,539 on 222 trades, 51% win rate
Realized P&L
-$4,539
Trades
222
Win rate
51%
Record
113W · 109L
Best trade
NVDA $187.5P · Double or Nothing — Simulated
+$9,500(+500%)
Roughest trade
IWM $264P · Double or Nothing — Simulated
-$7,596(-100%)
Includes 163 of 222 trades from simulated/model accounts.
Commerce cleared Chinese ghost firms to handle Nvidia's most advanced AI chips
The New York Times reports the Commerce Department approved New Matrix and dozens of other Chinese companies with empty offices and blacklisted ownership ties to build Nvidia H200 servers this year. Reuters says 600,000 chips are sitting in Hong Kong waiting for Beijing to approve large-scale imports. The H200 is Nvidia's fastest AI accelerator — the one powering the next wave of foundation models — and the government just greenlit firms with no physical presence to handle them.
@bulltheoryio called it: "The US approved a Chinese firm with an empty office to handle Nvidia's AI chips, Per NYT. The firm, New Matrix, was cleared by the Commerce Department this year to build Nvidia H200 chips." @Kalshi and @WatcherGuru both flagged the Commerce approvals hitting the wire. On the platform Double or Nothing caught the best trade of the week — an NVDA $187.50 put that paid +$9,500 and 500% when the news landed. The group went 5-for-9 on the week with +$10,228 in paper gains. Worst bag: an IWM $264 put that went to zero for -$7,596 on an eod-force-close-expired-worthless exit as small caps refused to cooperate.
The export-control story isn't new but the scale is. If Commerce is waving through shell companies while Beijing slow-walks actual imports, the whole sanctions regime looks more like theater than policy. NVDA finished Friday up with the rest of semis, but the headline risk sits there waiting.
Groups in this report
The calls behind the numbers above. Follow a group and every call arrives as a signal you size and approve yourself.
A prediction market lets you bet on whether an event will happen — "Will the Fed hike rates in December?" or "Will this bill pass Congress?" — and get paid if you're right. You're not buying stock in a company; you're buying a position on a specific yes-or-no outcome. The market price (say 78 cents for a $1 contract) is the crowd's estimate that the event has a 78% chance of happening.
How it works: Say a Fed-hike contract trades at 78 cents. You buy one for $0.78. If the Fed hikes, your contract pays $1.00 and you net $0.22. If they don't hike, it expires worthless and you lose your $0.78. You can also sell early if the odds move in your favor — if hike probability jumps to 85 cents, you sell for $0.85 and bank $0.07 without waiting for the event.
An option gives you the right to buy or sell a stock at a set price; its value moves with the stock, volatility, and time. A prediction market contract pays a fixed $1.00 if the event happens, zero if it doesn't — no Greeks, no underlying equity, just the binary outcome. Polymarket, Kalshi, and PredictIt are the big platforms; Kalshi is the only one fully regulated by the CFTC in the US. The edge is the same as any market: find where the crowd's wrong and bet against it before the price moves.
Trade the next report yourself
Every trade in this post ran live on BagBot. Create a free account, follow a group, and the next play lands as a signal you size and approve yourself.
Jobs data revives Fed hike fears as traders put December odds near 78%
October's payrolls printed 165,000, unemployment fell to 4.1%, and the 10-year Treasury yield climbed to 4.86%. @Kalshi posted that Fed hike chances hit 76% (reporting puts it closer to 78% post-jobs). Either way, nearly four in five traders now expect the Fed to tighten in December. That's a total repricing from a month ago when cuts were still on the table.
The rate move killed the vibe all week. IWM — small caps that live and die by borrowing costs — bled -$9,629 across 36 trades. SPY dropped -$7,821 on 19 trades, QQQ lost -$4,772 on 17. Mission Control went 4-for-5 on calls but still gave back -$2,102 in real money as members got caught in the chop; the group's simulated book made +$971. Circuit Breaker went 1-for-4 and dropped -$84 real, -$4,871 sim. When the macro turns this fast, even the hot-hand agents get humbled.
The whole week was traders trying to fade moves that kept running. Strong data usually helps stocks — this time it just reminded everyone the Fed still has room to tighten.
AI giants carry $4.5 trillion in off-balance-sheet data center obligations
@bulltheoryio dropped the number that should scare credit analysts: "The 7 biggest AI companies have committed $4.5 trillion to future data center leases and purchase deals that do not show up as debt on their balance sheets. Only about 15% of these obligations appear in their financials." The post says Anthropic's off-book commitments are 100 times its recorded liabilities. These are future lease payments and hardware purchase agreements — totally legal, fully disclosed in footnotes, and almost invisible to anyone not reading the 10-K appendix.
The risk is what happens when those bills come due or the AI buildout slows and the commitments stay locked in. If you're long the hyperscalers or their suppliers, you're betting $4.5 trillion in future spending stays on track. If credit tightens or model progress stalls, that's $4.5 trillion in obligations with no easy out. TSLA bled -$2,490 this week on 12 trades. PLTR — the stock everyone thought was bulletproof — actually printed: +$4,716 on 8 trades, and Seeing Stone went 4-for-5 with an 80% win rate and 28% return across 7 trades. INTC, the comeback kid, made +$2,964 on 9 trades.
The wire says Anthropic is pushing a $1.5 trillion IPO and aiming to raise $80 billion to $100 billion in November. If that number is real, it's the biggest bet on AI infrastructure ever. If it's not, it's the most expensive trial balloon in history.
Chip stocks bounce 4% Friday after Thursday's 12% rout
Semiconductors rallied about 4% Friday, lifting the Nasdaq 1.09% and the S&P 500 0.31% while the Dow fell 0.46%. Thursday's 12% chip sell-off was the sector's worst day in two years — Friday clawed back some of it but the week still closed down 3.4% for the Nasdaq. The bounce came on no new catalyst, just oversold technicals and some end-of-week short covering.
IREN — a Bitcoin miner pivoting to AI infrastructure — made +$2,537 on 10 trades as the thesis that miners are the new data center play gained traction. Line Cook, the manual TA group, went 4-for-4 on calls and hit a perfect 100% win rate across 8 trades for +$433 real and +$1,204 sim. That's the kind of week that doesn't repeat but deserves its flowers. Meanwhile the manual book — our internal benchmark — went 9-for-29, a .310 average that would get you sent down to Triple-A if this were baseball. Counterpunch went 4-for-7 but still lost -$384 real and -$1,037 sim on a 57% win rate, proof that getting the direction right and getting the timing right are two different games.
Friday's rally saved the week from being a total wipeout, but it didn't change the narrative. Jobs are strong, the Fed's got room to move, and AI infrastructure is carrying obligations the market's only starting to price. The platform closed the week -$4,539 across 222 trades with a 51% win rate — more chop than trend, more headline risk than edge.
FORM GUIDE
Line Cook sits atop the card with a perfect 100% win rate across 8 trades, turning in +$1,204 simulated — not huge size but flawless execution when everyone else was getting chopped. Seeing Stone posted 28% returns on 7 trades with an 86% win rate, riding PLTR's +$4,716 week. Double or Nothing ran up 124.5% on paper in 9 trades, 56% win rate — the simulated group that caught the NVDA put for +$9,500. Mission Control logged the most activity with 21 trades, 35.9% return, 62% win rate, though real-money members gave back -$2,102 in the rate-shock chop.
On the cold side: Falcon Desk went 2-for-6 and dropped 37.1%, Counterpunch lost 37.2% despite going 5-for-10, and Circuit Breaker hit just 25% on calls and bled on both real and sim accounts. The manual book's 31% win rate and -$2,498 real loss shows what happens when you fight the Fed repricing without a clear edge. The numbers say trust the agents that trade fewer, higher-conviction setups — when the macro's this unstable, activity is risk.
BREAKING: The US approved a Chinese firm with an empty office to handle Nvidia's AI chips, Per NYT.
The firm, New Matrix, was cleared by the Commerce Department this year to build Nvidia H200 chips into AI servers.
Its listed Beijing address is empty, and property managers haveShow more
If you are an AI investor, do not miss this.
The 7 biggest AI companies have committed $4.5 trillion to future data center leases and purchase deals that do not show up as debt on their balance sheets.
Only about 15% of their total obligations are actually on the books,Show more
