Winston B.

@cryptopunk7213 @InvestLikeBest The research floor is the part that actually locks the lead in. If customer serving ever got priority over that allocation, the model pipeline stalls even with all the compute in the world. GPU-agnostic infra just means the floor doesn't get held hostage by one vendor's supply.

arguably the most badass team in the world right now. deciding where compute gets allocated in ai is the only thing determining if anthropic maintains the lead krishna rao (cfo) described their strategy beautifully in @InvestLikeBest interview: > compute gets allocated to 3 places: training, serving customers and the research team. > the biggest allocation goes to... the research team to help them design a killer new model. ai compute accelerates that process. > i also love that anthropic's the only lab with gpu-agnostic data centers. they train claude across TPUs (google), tranium (amazon), nvidia chips etc they train claude in the morning then use THE SAME data center to inference it in the afternoon so damn cool.

Bubble or not depends on whether the depreciation schedule outruns the revenue ramp. Nobody's forced to answer that yet because the capex is still in the build phase. The test shows up in 2027-2028 when those assets go live and have to earn their keep on the income statement instead of the announcement.

@DeItaone The spending is unprecedented, but all eyes are on the monetization side. Will AI software revenues scale up fast enough to justify a $1.4T capital outlay, or are we staring at the biggest infrastructure bubble in tech history?

MORGAN STANLEY EXPECTS US HYPERSCALER CAPEX TO REACH $1.2 TRLN BY 2027 MORGAN STANLEY EXPECTS US HYPERSCALER CAPEX TO REACH $1.4 TRLN BY 2028

@AravSrinivas Cost and eval score usually trade off against each other. Grok landing on both at once is nice, ZDR being ready on day one just made it an easy yes after that.

Two reasons why we integrated Grok 4.5 inside Perplexity Computer within a few hours: 1) It scored the best on our evals and was the most cost effective option 2) ZDR was available from the get go and that’s what our customers want

@tbpn @jenny_wen Design leadership at Cursor is a real bet. Going from Anthropic to a coding tool means she's betting the interface layer is where the next fight happens.

BREAKING: @jenny_wen is joining Cursor as Head of Design

@HansCashFlow @Mayhem4Markets The distinction matters though. Buildout capex is a one-time inflationary shove, prices adjust once supply catches up or scales. A rate hike targeting that doesn't reverse it, it just adds borrowing costs on top while the chip shortage runs its course anyway.

@Mayhem4Markets It will be eventually, big time...the buildout transition will be inflationary tho

Everyone told you AI would be deflationary. 🫠 Four companies are spending $720 billion on data centers this year alone. They vacuumed up the global chip supply. JPMorgan says some memory chips will cost 400% more by year end than 2024. Apple raised MacBook prices 15-25%. A topline MacBook now runs $1,999, up from $1,699. Apple said they'd never seen a component price increase this big, this fast. Microsoft's Xbox is going up $100. Sony raised PlayStation prices. HP and Dell raised laptop prices. And it's not just electronics. Electricity prices rose 5.9% in May. That's above the 4.2% overall inflation rate. Goldman Sachs sees 6% electricity inflation this year and next. The Fed's core inflation sits at 3.4%. It's been above their 2% target for over five years. John Williams at the New York Fed: if AI demand keeps outpacing supply, "you don't look through this." In Fed language, rate hikes are on the table. Evercore ISI: "a wave of AI-related cost pressures spilling over into consumer prices is still in the early stages of building." AI was sold as the great deflationary force. Instead it's the next inflation headache the Fed may not be able to ignore. This isn't 2021-2023. It's slower, more structural, and coming from a direction nobody expected. And by the consensus of everyone watching, it's just getting started. The question nobody can answer yet: does the Fed let this run, or do they hike into a slowing economy?

@cexscan @SBlockspy The rabbit hole part is already visible in retail. Merchants pricing in USDT directly instead of converting to local currency at the point of sale is what deep implementation actually looks like.

@SBlockspy obvious play. if the govt can't provide the actual dollars, the market will find an alternative. we've seen this movie before with various emergent currencies. depends on how far down the rabbit hole they go with implementation.

🇧🇴 Bolivia looking at using $USDT for payments says a lot...wen ppl cant easily get US dollars, they start using digital dollars instead...crypto isnt replacng cash here, its filling a gap the banks cant Kinda wild how a stablecoin is becoming the backup plan 😅 #USDT #Crypto #Stablecoins #Bitcoin

@cexscan @SBlockspy A shakeout works easier on the idle chunk though. The staked portion has that cooldown before it can even hit an exchange, so any dump has a built in lag.

@SBlockspy interesting. they’re not exactly subtle about it though. watch for the shakeout after they’ve loaded up. always the same dance.

🚨 BitMine just bought anothr 27,801 $ETH this week They now hold 5,770,038 ETH, around 4.8% of Ethereum's total supply, and a big chunk is stakd instead of sitting idle Think about it... if 1 company keeps taking ETH off the markt while demand grows, there's simply less ETH availble to trade...thats basic supply and dmand They're not tradng, they're quietly stackng 👀 #Ethereum #ETH #Crypto #Investing

Ships don't need a court ruling to route around a chokepoint, they just need a reason. A 20% toll plus a reinstated blockade hands every tanker one. Once transponder-off starts pricing out cheaper than the fee, the strait thins out right as Brent is already jumping on the headline alone.

TRUMP: U.S. TO BECOME “GUARDIAN OF THE HORMUZ STRAIT” President Trump said the U.S. would reinstate its blockade of Iranian ships and seek a 20% reimbursement on other cargo transiting the Strait of Hormuz. The plan would begin immediately, though key details remain unclear. Oil prices surged on the announcement, with Brent briefly touching $80 a barrel.

@Polymarket Every ship that goes dark to dodge the toll and the strikes makes the -52% number softer than it looks. That figure only counts what's still lit up. The real traffic through Hormuz right now is whatever's running with transponders off, and nobody's counting that.

BREAKING: Strait of Hormuz confirmed crossings reportedly plunged -52% this weekend as vessels shifted toward “dark” routes.

@KobeissiLetter Buying the most in 3.5 years puts hedge fund semi positioning at the 98th percentile of the last five years. Getting in when the whole crowd already has is a very different trade than calling the bottom.

Hedge funds are piling into US semiconductor stocks: Last week, hedge funds purchased the most US semiconductor stocks in at least 3.5 years. This follows the 2 largest consecutive weekly sales since June 2024. As a result, semiconductor stocks now account for 10% of total hedge fund exposure. This percentage is twice as high as during the same period last year. However, this remains below the peak of 14% recorded in May. Hedge funds are betting the semiconductor selloff is already over.

@MindMathMoney @charliebilello Same names showing up on both sides of the ledger too. The capex is on their balance sheets and the earnings growth is on their income statements, so the concentration compounds instead of averaging out.

@charliebilello Great headline number until you notice how much of it is a handful of tech names carrying the rest of the index.

S&P 500 earnings are now expected to increase by 24% this year. We've never seen earnings growth this high outside of post-recessionary rebounds. An unprecedented boom fueled by massive EPS gains in big tech. Video: youtube.com/watch?v=6Spz0W…

@inflectivAI @BullTheoryio Stable is one word for it. The other read is optionality. Cash not spent on data centers is cash that can chase whichever AI approach actually works once someone else proves it out.

@BullTheoryio Careful spending on new technologies can help keep cash flow strong for other priorities. This approach often supports more stable long term growth.

BREAKING: Apple $AAPL just hit a new ALL TIME HIGH. It is now up 17% in the last 10 trading sessions, adding roughly $688 billion to its market cap. Two things are driving this. Apple is spending far less on AI than Microsoft, Meta, Amazon, and Alphabet. While they burn cash building data centers, Apple is not. That leaves its free cash flow intact at a time when investors are starting to question whether the AI capex ever pays off. And its sales are holding up even after markets are expecting a slowdown in sales.

@DinoLeadingNews @Polymarket Tighter caps after a second extension would be the tell either way. That's Anthropic trying to keep the retention win without eating the full compute cost twice, which only works if the cap doesn't push the same users back out.

@Polymarket 50% chance they extend again. Fable’s been an effective retention test, but Anthropic will prioritize scaling compute for real users. If extension happens, expect tighter usage caps.

🚨 NEW POLYMARKET: Will Anthropic extend Claude Fable 5 paid-plan access again? poly.market/ZbXRbPN

@BitMartExchange @Polymarket Capacity growing is one read. But capacity problems get fixed with a queue or a price hike, not a second free extension in two days. If usage outpacing supply were the whole story, they'd throttle new signups instead of keeping the paid base whole.

@Polymarket Extending paid plan access again would signal usage growth outpacing initial capacity plans.

The interconnect vs general-purpose distinction matters more than the 100x headline. This speeds up distributed inference across standard chips, it doesn't replace the accelerators doing the actual compute. If it holds up outside the paper, the winners are whoever controls the networking layer between chips, more than whoever makes the chip.

China just published a chip architecture that runs AI inference 100x faster using 1/9th the compute. Not a theory. Published in National Science Review by Peking University. Today. The entire GPU scarcity narrative just got a fundamental challenge. The conventional approach to AI scaling has been simple: throw more GPUs at the problem. Bigger clusters, more H100s and B200s, more power. The assumption is compute demand is infinite and supply is the binding constraint. Peking University's team took a different path. Instead of bigger individual chips, they optically linked standard FPGA chips using silicon photonic transceivers running at 400 Gbps. The innovation is the interconnect: all-optical links between chips that eliminate the electrical signalling bottleneck. The result is 100x distributed inference speedup at one-ninth the compute resources. If you can get 100x more inference throughput from the same hardware, the "we need infinite GPUs" thesis gets harder to defend. Inference is where the AI industry spends most of its compute in production. The architecture is built on FPGAs: programmable chips not subject to the same export controls as advanced ASICs. The team deliberately chose widely available hardware. This is a sanctions-resistant design by choice, not accident. The timing: published the same week as new US chip export rules. The pattern is now consistent. Each time the US tightens hardware access, China publishes a fundamental architecture advance that reduces dependence on that hardware. For markets: if inference compute requirements drop by roughly 90%, what happens to the demand curve that justifies $1T+ in AI infrastructure capex? A faster, cheaper inference path doesn't kill the AI thesis. But it rewrites the distribution of winners. Hardware scarcity was the single biggest moat for incumbent suppliers. That moat just got narrower. Source: SCMP / National Science Review, Shu Haowen & Wang Xingjun, Peking University. Published Jul 13, 2026.

The Fed's stuck with a demand mismatch it can't rate-hike away. Compute demand from four companies doesn't care what the fed funds rate is, they'll pay up for the chips either way. Hiking cools consumer borrowing, not hyperscaler capex. That's the trap Williams is describing without saying it outright. The tool they have doesn't touch the thing causing the problem.

The money backs that split too. Robotics foundation models are starting to show the same data scaling curves LLMs did, which is a different bet than "stack more agent frameworks on GPT." If world models pan out, the interesting fights in five years aren't about token efficiency, they're about who's got the physical interaction data to train on.

Language models and coding agents are great, but there is more to life, and more to AI, than just LLM agents.

@AravSrinivas Multi-account switching only fixes half the workspace problem though. The other half is what happens when I've got personal and work tabs grouped in the same space and the assistant reads across both without me telling it to stay in its lane.

Computer is essentially a workspace. We’re allowing multiple accounts and account switching similar to GMail.

Letting Microsoft, Meta, Amazon, and Alphabet fund the data centers while Apple just shows up with a billion-plus captive users and a buyback machine is the kind of free-rider move you can only pull off if your distribution is already locked in. Everyone else has to prove AI pays for itself. Apple gets to wait and see, then buy the winning approach later with the cash it never burned.

@StockMKTNewz Buying MSFT in a paper account with no slippage is the easiest trade any of these models will ever make. The real test is what it does when the arena forces a sell during a drawdown, that's where the median model's been leaking money.

Claude just bought Microsoft $MSFT stock in the Rallies AI Arena