Tale of two copper markets rn. West dumps it, China vacuums it all up.
Yangshan premiums are mooning and SHFE literally pushed expiring holders into cash settlement bc they're out of deliverable copper.
China is super tight while Western funds panic sell over AI bubble & recession fears.
#copper
Here's some @TomorrowXSummit Day 1 takes for anyone that cares - The variety of conversations across public pensions, credit, long/short funds, power and compute markets has been useful for seeing how differently people underwrite the AI buildout.
• The strongest private skeptical argument was token output improving faster than new uses can absorb it. Rental prices and GPU residual values become difficult to underwrite without another wave of applications.
• Credit investors focused on operator quality and guarantees. Power investors liked specific contracted projects. Conviction in individual financing structures was stronger than conviction in returns across the whole buildout.
• Compute traders saw older hardware becoming more interchangeable. Neocloud participants pushed back on memory, precision and networking requirements. Those differences complicate trading and hedging compute.
• Public-equity investors saw industrial companies benefiting from AI cost savings while still grappling with expectations for select stocks as growth outlook is more mixed.
Main Stage takes - mostly hosted by Gavin and @AntonioGracias@GavinSBaker / Atreides: His roughly 60-year railroad analogy depends on demand remaining economically useful. Higher rates and power/wafer shortages could restrain overbuilding and extend the investment cycle.
@altcap / Altimeter: Cited Anthropic approaching a $100B revenue run rate with roughly 5,000 employees, versus Google’s 105,000 at that run rate. Estimated inference used roughly 500 MW of its 2 GW. His argument was that revenue can grow rapidly without proportional growth in inference requirements.
@DavidSolomon / Goldman Sachs: Administrative AI is scaling. Goldman expanded from six operating processes under redesign to eight, including onboarding and anti-money-laundering checks. Savings can fund additional client coverage.
@LisaSu / AMD: Bullwhip risk as shortage-driven orders become excess inventory. Different workloads need CPUs, GPUs and custom silicon. Her example of an engineer using Claude to improve an AMD implementation showed how software optimization can make alternative hardware more competitive.
Stan Druckenmiller / Duquesne: “I’m not playing.” Short copper, roughly flat net equities. Infrastructure looks more like a tactical trade after its prior run. His longer-term interest is AI use cases, including labs moving from selling tokens to selling completed tasks.
@MichaelDell / Dell Technologies: No change to his three-year buildout view. Roughly 6,500 AI Factory customers provide demand beyond frontier labs. His September update included 3,300 added in three quarters. Private deployments make API revenue an incomplete adoption measure.
@DinaPowellMcC / Meta: Local tax receipts are the political argument. Her Louisiana example was teacher bonuses rising from $10,000 to over $50,000 through data-center tax revenue. The case for community support gets stronger when residents can see the buildout in their own paychecks.
Jitendra Mohan / Astera Labs: Customers asking for more compute over the preceding two weeks, with Baker later referencing requests for earlier deliveries. Expansion into switches, memory connectivity and optics follows copper’s physical limits.
@rauchg / Vercel: $700M annualized revenue, up 180% year over year. Ten years to its first billion deployments, now 1.25B every four months. Agents expand both the builder population and output per builder.
@lqiao / Fireworks: Still seeking compute. Autonomous workflows remove the human pause between prompts = explosion in future usage. Running software on customer and partner capacity could let growth exceed Fireworks’ ability to finance hardware.
Bret Johnsen / SpaceX: Put early terrestrial builds in the low-to-mid $30s/W versus $50–60/W for others, in his comparison. Power generation, vertical integration and faster construction underpin the advantage. Aims to be among the largest AI infrastructure providers in 2027. His orbital argument is that competitors face rising power, labor and construction costs while reusable launch drives SpaceX down its cost curve. As chips get cheaper, those non-chip costs matter more. Targets meaningful orbital gigawatts in 2029.
Marc Rowan / Apollo: His dollar of credit needs to come back with its contractual return. Equity investors need future compute earnings to make their dollar worth more. He questioned whether investors throughout the stack are being paid adequately for the risks they’re taking. Discussed amortizing chips to zero over five years and financing against cash flows and counterparties he can underwrite today. Disappointing AI revenue can strand assets and impair returns while the hyperscalers themselves remain solvent.
@alighodsi / Databricks: Agents underappreciated, recursive self-improvement overhyped. Adoption slows as industries become more regulated, leaving existing capability underused. Even inside Databricks, he had told employees the prior week to ask the model when it knew more than they did. The opportunity is getting organizations to use intelligence already available and making it less complicated.
@tbpn interviews w/ @jordihays@johncoogan@DavidSolomon: Thousands to tens of thousands of AI-assisted vulnerability scans daily for seven or eight months. Faster patching also creates operational risk.
@GavinSBaker: A handful of dominant labs could recreate concentrated cloud-buyer economics for semiconductor suppliers. Open models broaden the customer base.
@andrewdfeldman / Cerebras: Uses 5 nm and avoids HBM to compete with Nvidia for fewer scarce inputs. Sees opportunities for 50–200 MW inference sites and expects roughly 75% of Cerebras’s 2027 data centers outside the U.S.
@ChaseLochmiller / Crusoe: Power availability determines where compute goes. Crusoe Spark puts prefabricated compute beside that power. Factory construction reduces dependence on scarce local construction crews, and Crusoe makes its own electrical switchgear. The strategy is to control more of what determines how quickly available power becomes operating compute.
There are 14 U.S. warships forward-deployed to Yokosuka, Japan, the tip of the spear of American deterrence in East Asia.
Six of them are now in the Middle East and the Indian Ocean.
USS George Washington (CVN-73)
USS Shoup (DDG-86)
USS Milius (DDG-69)
USS Rafael Peralta (DDG-115)
USS John Finn (DDG-113)
USS Mustin (DDG-89)
news.usni.org/2026/11/16/usn…
There are 14 U.S. warships forward-deployed to Yokosuka, Japan, the tip of the spear of American deterrence in East Asia.
Six of them are now in the Middle East and the Indian Ocean.
USS George Washington (CVN-73)
USS Shoup (DDG-86)
USS Milius (DDG-69)
USS Rafael Peralta (DDG-115)
USS John Finn (DDG-113)
USS Mustin (DDG-89)
news.usni.org/2026/11/16/usn…
World leaders are heading to Shenzhen for APEC. The robots are already on shift!
At the APEC media center, robots from AI² Robotics are making and serving coffee, ice cream and juice for journalists. Robots from X Square Robot are serving drinks and snacks and handling cleaning duties.
Qianhai Stone Park is one of the APEC tour stops, and the volunteers there have robot coworkers helping with directions and handing out supplies.
At Bao'an Airport, robots are rounding up luggage carts, cleaning the floors and patrolling alongside police. You can even order a drink on your phone and have a robot bring it to you.
Shenzhen is also hosting a robotics carnival during APEC. More than 100 robotics companies and suppliers will bring hundreds of robots for live demos and hands-on experiences.
The Lutnick clan even more "lucky" than the 'Larry Silverstein' clan.
Get over it you are a born loser, he is a winner, don't be anti-septic just cause you lack his talent.
Some coincidences are hard to explain.
The official who controls Nvidia’s chip exports to China just saw his family firm dodge Nvidia’s crash.
Commerce Secretary Lutnick’s Cantor Fitzgerald cut its $NVDA stake by -45% between July and September, per its latest 13F.
It also held put options on 1.3M $NVDA shares, positions that profit when the stock falls.
Weeks later, $NVDA peaked, then crashed -31% in just 6 weeks.
So they sold before the crash and loaded puts that paid off on the way down.
And during those same months the firm was selling, Commerce was refusing to act on suspected Nvidia chip smuggling to China.
Treasury, Defense and State pushed for a crackdown, and Trump ordered one 3 times this year, but Lutnick reportedly stalled each one, per Axios.
A blacklist could have hit Nvidia’s sales and sent the stock lower while Cantor was still selling.
Instead, the Nvidia scandal only blew up in November, after the firm had already cut its stake.
Talk about timing.
Most robot policies (VLAs) are built on vision language models trained on static web images. It doesn't have the physics understanding of the environment.
Mimic video takes a different route. It’s a Video-Action Model (VAM) that grounds the policy in a pretrained video model, so the video backbone handles the physical dynamics and the policy only has to handle low-level control.
Unless you’re still clinging to Jevons paradox, which has yet to show up in any meaningful way, this is bearish for AI infrastructure stocks. More intelligence for less money is great for customers, but not so great for infrastructure providers counting on explosive revenue growth from frontier models to cover their long-term commitments. Compute glut.