AI token prices hit a record low while Dell guides to $192B partly on cost pass-through. Two labs shipped cheaper models the same day. OpenAI flagged its own upcoming model as a cyber risk. Barr and Bessent publicly disagreed on September.
Futures Are Flat. Asia Sold Off Hard. WTI Crossed $90.
S&P 500 futures are barely changed after three straight losing sessions. Asia took the brunt overnight. Japan's Nikkei fell sharply. South Korea's Kospi dropped hard. WTI crossed $90, its highest since late July, on continued U.S. strikes against Iran. The global bond selloff deepened. The 10-year pushed higher again.
Broadcom (AVCO), Hewlett Packard Enterprise (HPE), and Snowflake (SNOW) report after the close. ADP private payrolls land before the open. The Fed Beige Book publishes later. Dell (DELL) jumped sharply after hours after raising its full-year revenue guide by roughly $25 billion.
Investor Signal
Three losing sessions in a row. Asia down sharply. Oil at a two-month high. Futures flat says the market is not panicking, but it is not adding either. ADP and Broadcom are the two prints that set the tone from here into payrolls Friday.
In January, Gold Touched Nearly $5,600 an Ounce. Today It's Around $4,100.
So the story's over, right?
Then explain this.
The metal is still leaving the vaults. Physical deliveries still running at levels the exchange rarely processed before. Central banks still buying. Dealers charging 30-40% premiums over paper price for real coins.
When price falls but physical demand doesn't — only one of those two is telling the truth.
The paper market sets the price. The physical market sets the deadline.
Anyone who wished they'd bought miners before January's run just got handed the entry back.
One company I've been tracking controls an 88 million ounce deposit — trading near $4 billion. About 1% of the value of its metal in the ground.
That gap is the whole opportunity.
Token Prices Hit a Record Low. AI Is Deflating What It Sells While Inflating What It Buys.
Silicon Data's LLM Token Expenditure Index fell to 97 cents. That is the lowest reading since the index launched and more than half below its summer high. Open-source Chinese models, OpenAI price cuts, and dynamic pricing all pushed the rate down.
The compression is hitting a fixed cost base. Anthropic's $45 billion Nscale commitment and its $35 billion Lambda deal are priced in dollars. The tokens those data centers produce are now worth less than half what they were a few months ago.
Cost goes up. Output price goes down. The gap is the margin question nobody has solved yet.
Deflation Has a Timeline
- Token prices at 97 cents arrived weeks before Anthropic's IPO
- Enterprise pricing in the S-1 is the first public disclosure of what labs actually charge
- The index may stabilize above 90 cents or keep falling through the listing window
Token deflation is not a demand problem. Demand is growing. It is a pricing power problem. And Anthropic's prospectus will be the first document that has to address both at the same time.
Cost Moving Up, Price Moving Down
Dell is passing memory costs forward into a $192 billion guide. Apple raised MacBook and iPad prices for the same reason. At the same time, the output of the systems those chips power is falling in price. Someone in the stack absorbs the gap. The question is who.
Dell Raised Its Full-Year Guide by $25 Billion and Said Price Increases Are Inside It.
Dell beat on both lines and guided fiscal 2027 revenue to $192 billion against a consensus of roughly $173 billion. That is a $25 billion raise from the guidance it gave in May. Operating chief Jeff Clarke confirmed price increases from climbing input costs are built into the elevated number.
AI-optimized servers drove the beat. Traditional server and networking revenue jumped sharply. Clarke said customers are deploying incremental CPU capacity to support agentic workloads. That is new demand on top of AI infrastructure spending, not a replacement for it.
Where the Beat Lives
- AI server sales guidance nearly tripled from six months ago
- Traditional server growth accelerated sharply, driven by agentic workloads
- Client solutions came in slightly below consensus despite strong year-over-year growth
- Price increases from memory and component costs are embedded in the $192 billion
The $192 billion guide depends on both demand holding and input costs not rising further. Broadcom tonight is the next read on whether the demand assumption holds.
Margin Is the Question
Clarke named the mechanism. Component costs rise. Dell passes them forward. Hyperscalers receive them. They sell tokens into a market where the price just hit its lowest point on record. The chain has to break somewhere.
AI CEO Issues Code Red: Prepare for Meltdown
The CEO of this AI company (click here to get the name, 100% free) just issued a CODE RED in an internal memo…
Warning his employees that they’re dealing with a critical situation.
Another company executive even implied they might need a government bailout.
And now Jim Rickards is predicting this company is about to go bust, in a full-blown AI meltdown that could be 10 times bigger than Lehman Brothers.
Two Labs Shipped Cheaper Models the Same Day. Google Scrapped Its Flagship.
Anthropic released Fable and Mythos 5.1 with reduced token cost as a headline feature. The system card is unusually direct: the models cooperate with human misuse and accept unverifiable claims somewhat more readily than the prior version. Anthropic is shipping less restriction alongside lower cost.
Google is releasing Gemini 3.8 Flash, a smaller and cheaper model. Its Pro series is months behind schedule. The internal Pro candidate was scrapped because it was not sufficiently better than the cheap Flash model. That is not a strategic choice. It is an acknowledgment that the cheap model is now the benchmark.
What the Scrapped Pro Says
Google's internal 3.5 Pro failing to beat Flash is the same conclusion the token index reached independently. For most tasks, the cheap model is already good enough. The moat around frontier capability is now measured in months, not years.
OpenAI Rated Its Own Upcoming Model a Critical Cyber Risk.
OpenAI's internal testing found its forthcoming Astra model can devise and execute novel cyberattacks with limited human input. It compromised browser sandboxes and found vulnerabilities in a hard-to-hack operating system. OpenAI rated it "critical" for cybersecurity in its internal risk framework. That is a first.
The response: extra training to refuse malicious requests, restricted public release, and a small initial tester group. OpenAI is shipping less of its most capable model because of what it can do.
Palo Alto Networks (PANW) grew revenue sharply and posted a net loss after acquiring CyberArk and Chronosphere. Shares fell after hours despite the beat.
Same Demand Curve, Two Sides
OpenAI is restricting what it ships. Palo Alto is selling remediation for what gets out anyway. The customer pays for both. CrowdStrike (CRWD) is in the same position. Security demand grows directly from frontier capability risk. That loop is not going to close.
Where the Revenue Goes
Palo Alto's 34 percent revenue growth into a net loss shows what buying the thesis costs. The CyberArk and Chronosphere acquisitions were bets on consolidating the security stack before AI rewrites it. Whether those bets clear the hurdle rate depends on whether OpenAI's restrictions hold.
On September 8th, a powerful new law signed by President Trump
will trigger a radical shift in America’s money system...
When a small group of private companies — not the Fed — will perform a major mint of a new kind of money.
And those who act before this new system fully kicks in could see gains as high as 40X by 2032.
But those who fail to prepare will be blindsided by this sea change to the U.S. dollar.
Go here now for the details — before the September 8th mint hits the market.
Barr Said Hike. Bessent Said Don't. They Both Have Data on Their Side.
Barr said the Fed should begin raising rates unless new data show price pressures easing. He is a permanent voter. Three colleagues already voted to hike in July.
Bessent pushed back publicly on CNBC. His argument: you don't raise rates into a supply shock unless second-order effects appear. He named core inflation as restrained. The ECB raised in June and is expected to raise again next week, treating energy inflation as a risk to expectations rather than a reason to wait.
Token prices are falling. Dell is raising prices. Core has held. Goods inflation has not. Bessent's supply-shock defense holds only if component pass-through stays out of the next print.
September 11 Is the Arbiter
CPI lands September 11. The meeting is September 15. Both sides are arguing from the same data. The only new input either gets is that print. Whoever it favors wins the argument. The market is already priced for a coin flip.
Wednesday opens with AI deflating its own output price while the hardware driving it passes costs forward. Dell raised its full-year guide by $25 billion with price increases embedded.
Two labs competed on cheap models the same day. OpenAI shipped less of its most powerful model because of what it can do. And Barr and Bessent publicly disagreed on September with the same data.