Chip stocks slide in wake of ‘casino’ comment by Broadcom boss

A global sell-off in semiconductor shares deepened on Tuesday as comments from Broadcom’s chief executive about the casino-like nature of AI investing knocked investor confidence in high-flying stocks ahead of the start of Wall Street’s earnings season.

Shares in Samsung Electronics and SK Hynix, the world’s two largest memory-chip makers that together account for about half of Korea’s Kospi index, fell 7.4 per cent and 10 per cent respectively. Futures tracking Wall Street’s Nasdaq 100 tech index were 0.4 per cent lower after a fall in semiconductor stocks dragged the index down 0.7 per cent on Monday.

Concern about the sector was fuelled by Broadcom boss Hock Tan’s warning about the gambling-like characteristics of the AI trade. “A lot of investors have no clue what they are investing in,” the head of the US chipmaker told an event in Singapore on Monday in comments reported by Bloomberg. They “set very extreme expectations only to find out that it’s almost like they are in a casino”, he added. (…)

Adding to investor nerves, Chinese DRam producer CXMT claimed that it had made progress in producing cutting-edge DRam technology that Korean memory giants SK Hynix and Samsung have also been developing. The reports reignited fears about low-cost alternatives upending the leading AI companies.

The Hefei-based chipmaker said it would release a new server memory chip based on the design, which involves fitting more cells on to a single chip. The announcement by CXMT President Cao Kanyu at an industry conference over the weekend, with little technical detail to accompany it, would mean, if confirmed, that the company has beaten larger rivals to a significant breakthrough despite not having access to the leading manufacturing equipment from Dutch supplier ASML. (…)

Broadcom’s Hock Tan directed his skepticism at investors backing frontier large language models, where, in his view, enthusiasm has outpaced any grasp of the underlying economics.

He questioned whether the AI industry is generating enough economic value to justify the enormous sums being invested.

Tan said many businesses have adopted AI but have yet to find applications that meaningfully improve productivity or deliver a worthwhile return on investment. “People are just wasting money, spending tokens, trying to understand (AI)”.

He compared businesses using AI without clear results to a child spending arcade tokens without learning anything. His concern is that companies may be consuming large amounts of AI computing services simply to experiment, rather than to make their businesses more profitable.

Tan also argued that access to capital could become the biggest bottleneck for AI expansion, as developers need enormous amounts of financing to build computing infrastructure and train increasingly powerful models. “The biggest constraint is capital,” he said.

He also suggested that the current hardware boom cannot sustain itself indefinitely if AI model developers fail to create valuable products and services. Chipmakers can profit from selling the infrastructure, but ultimately their customers need profitable businesses of their own.

Tan remained optimistic about near-term demand for computing capacity, while acknowledging that the longer-term outlook depends on adoption, technological progress and continued revenue growth.

Broadcom itself guides to roughly $58 billion in AI semiconductor revenue this fiscal year, doubling to $115 billion in fiscal 2027 and again to $230 billion in fiscal 2028, projections Tan reaffirmed as recently as September. He also said demand for compute infrastructure still exceeds supply, an imbalance he expects could persist through 2029 and into 2030.

Intel CEO Lip-Bu Tan, on the same panel said:

• The biggest challenge is “capital and the cost of capital, and then the other part is in the longer run, people want to see the ROI”.
• He doesn’t expect a major slowdown in the next two to three years, because more people are using AI and demand for computing stays strong.
• AI will have an even bigger impact than the Internet, with AI agents making work more efficient.
• Chinese AI models such as DeepSeek, Kimi and GLM cost less to run and are “good enough,” and that people are adopting them “very fast”.
• “I’m a strong believer of open source, that’s the best way to drive efficiency.”

In case you missed my October 8 post: AI Capital Crunch?

So far, AI racers have been willing to pay the rising costs to participate. For some of them (e.g. Anthropic, OpenAI) this is an existential race.

They are also willing to pay the increasing cost of capital but, much like for the physical issues, capital is not infinite, particularly when its cost keeps rising. Lenders will eventually balk when they start questioning the safety of their capital given the ever rising borrowers’ liabilities.

If capital gets scarce at the top of the chain, the whole chain slows down.

• Meta Says New Muse Model Hits AI Frontier After Aggressive Push

Meta Platforms Inc. released a long-awaited artificial intelligence model on Monday, a milestone that will test whether its multibillion-dollar push to rebuild its AI organization has improved its technology enough to compete among leaders.

Known as Muse Wonder, the new model is the company’s most capable since Meta Chief Executive Officer Mark Zuckerberg overhauled his AI operation through an aggressive hiring push that put new talent at the helm. The company claims the model, dubbed “Watermelon” internally, is a big leap over its own predecessors, with “frontier capabilities” that its leadership say put it on par with OpenAI’s ChatGPT and Anthropic PBC’s Claude when it comes to coding, or working through financial or legal questions. (…)

Wang and his Meta Superintelligence Labs team are already looking beyond Muse Wonder. He said MSL has begun work on “larger and more scaled models.” (…)

Putin Ignores That Energy ‘Cease-Fire’ Trump announces a truce between Russia and Ukraine. Russia immediately violates it.

The WSJ Editorial Board:

President Trump announced the truce Sunday on Truth Social, along with the warning “do not break it!” Within hours Vladimir Putin did. (…)

Ukraine refrained from striking Russian refineries overnight, Mr. Zelensky said Monday, and “we are ready to adhere to any energy truce, if Russia is willing to do the same.” Mr. Putin has repeatedly refused.

The Russian dictator is multiplying his aerial attacks on Ukrainian cities and civilians as he has run into difficulties on the battlefield in eastern Ukraine. The Russians have made “near-zero gains since March 2026” (…).

So Mr. Putin’s plan now is to “make civilian life in Ukraine unbearable such that Ukraine will capitulate to Russia’s demands,” ISW said this month. Mr. Putin is taking advantage of Ukraine’s shortage of Patriot missiles and functioning F-16 jets to hammer energy plants ahead of a long winter.

Mr. Trump is scapegoating Ukraine for rising diesel prices, and on Saturday he said Kyiv “better damn well stop” its strikes on Russian refineries. But is Ukraine supposed to sit back and take the Russian blitz? Oil and gas exports fund Mr. Putin’s military, and Russian energy production is a legitimate war target.

Mr. Trump made a major concession Friday with a pause on diesel sanctions that will allow Russian energy companies to rake in some $4.5 billion. Yet Mr. Putin is giving him nothing we can see in return. The Russian refusal to accept an energy cease-fire is the latest evidence that Mr. Trump’s courtship isn’t working.

Trump yesterday: “I think actually Putin wants to make a deal more than Zelensky.”

Starving Ukraine of military equipment, blaming Ukraine for US diesel prices, unilaterally announcing a cease-fire, funding Russia’s war: the art of the deal!

The FT:

(…) On the day that Trump announced the Russia deal, his son-in-law Jared Kushner and chief envoy Steve Witkoff held talks with Ukrainians in Florida. They threatened to block weapons sales and end US intelligence sharing if Ukraine continued to strike Russian refineries.

This would deprive Zelenskyy of the last useful bit of US support against Russia’s escalating assault.

In the week before Trump’s diesel deal, more than 100 Ukrainians, including at least 20 children, were killed by Russian missiles. Another US president would be pressuring Russia to stop killing innocents. Trump is instead giving them the means to kill more.

He is also giving Russia a green light to keep helping Iran, which it has been supplying with satellite and tracking data to target US bases and allied installations in the region. In a very real sense, therefore, Trump is assisting attacks on America. (…)

• If the Saudis Will Fight the Houthis, the U.S. Can Help

The same Editorial Board:

(…) Now comes the test for U.S. policy.

Asked Saturday about joining Saudi strikes against the Houthis, President Trump said, “We may. We’re going to look at it.” Yet the last thing he seems to want before midterm elections is a military escalation. That hesitation may be what Iran and its Houthi militia proxies, whom it arms and finances, are counting on. (…)

The region is watching what Washington will do when the cornerstone of its Gulf alliance system is under attack. (…)

Under direct attack, Riyadh now has to decide if it has the will to fight back. If it doesn’t, all the modernizing ambitions of Crown PrinceMohammed bin Salman (MBS) will suffer. Few will want to invest in an easy, repeat target.

MBS’s new Mecca pact partners, Turkey and Pakistan, don’t seem eager to help despite their large militaries. That leaves an appeal to the U.S. as usual. U.S. support doesn’t mean standing in for Saudi Arabia but helping it lead. Pushing MBS into another deal with the Houthis instead would lock in the militia’s gains in preparation for the next war.

The Houthi challenge may seem obscure, but the stakes are high for the Saudis, the Middle East and the U.S. It is now tied to Mr. Trump’s campaign against Iran and his hope for an expansion of the Abraham Accords between Israel and the Arab states that someday includes Saudi Arabia.

As it stands now, the Houthis have the upper hand.

“Test for US policy”? What policy?

Just “a little excursion” he said 7 months ago.

• Why the Houthis Are So Hard to Defeat

(…) Announcing the start of a full-scale military campaign against the Houthis, the Saudi-led force said on social media on Oct. 4 that the goal is to “eliminate” the group.

It’s an ambitious aim. Once dismissed as a ragtag band fighting in sandals from pickup trucks, the Houthis have demonstrated staying power and reach. Having taken over the capital Sanaa in 2014, the Houthis now control land containing most of Yemen’s population. (…)

While the Houthis have long received military assistance from Iran, they have found other ways to arm themselves and shown growing military sophistication. (…)

The Houthis coordinate closely with Iran and largely share its strategic objectives, but they make their own decisions. Iran is an important source of weapons for the Houthis, but not the only one. (…)

The Saudi government has sought help from allies to counter the Houthis. The response has been tepid, suggesting a lack of enthusiasm to invest seriously in a fight against such a resilient foe. (…)

Meanwhile:

Supertankers Turn From Gulf Oil Transfer Hub Amid Iran’s Attacks

A growing number of empty supertankers have been turning away from the Persian Gulf as escalating Iranian attacks threaten a critical transfer hub for wartime oil shipments.

At least 10 very large crude carriers and several Suezmax vessels approaching the port of Fujairah in the United Arab Emirates have stopped or reversed course in recent days, according to tanker-tracking data compiled by Bloomberg.

While other vessels are still continuing on their journeys into the region, the retreat of these tankers amid near-daily attacks around the Strait of Hormuz — some near Fujairah itself — raises fresh concerns over the sustainability of oil flows through the waterway.

Fujairah is commonly signaled as an interim destination by vessels ultimately sailing to the Gulf of Oman — where they can collect cargoes via ship-to-ship transfer — or onward into the Persian Gulf. That makes the number of ships heading toward the port a broader indicator of tanker operators’ appetite for approaching the waterway.

The shift threatens a key workaround that has helped Persian Gulf oil exports recover to near prewar levels. Large crude carriers can wait off Fujairah and in Omani waters outside Hormuz while shuttle tankers make the riskier transit through the strait, bringing out crude and transferring it for onward shipment to buyers.

The system depends on a steady supply of receiving tankers, so a significant retreat from the region by empty VLCCs could slow the flow of crude, tying up the shuttle vessels for longer when tanker availability is already tight. (…)

• Kpler Suspects Gulf Producers Are Paying Iran for Safe Passage

(…) Iran could be quietly collecting tolls from ships transiting through the Strait of Hormuz, with exporters passing on these costs to consumers. According to Michelle Brohard, head of policy and geopolitical risk at Kpler, certain Gulf nations could have quietly struck deals with Iran to keep their oil flowing on fears that they cannot rely indefinitely on Washington to keep Tehran at bay.

“I suspect there is a toll that’s being paid, which is giving these ships safe passage,” Brohard said in an interview with energy analyst Rory Johnston last week. “I also suspect that these countries know that this is unsustainable from a perspective of [the] US escorting [ships], and also unsustainable from them paying Iran 10 percent of their cargo, or 20 percent of their cargo. So you’re starting to see like what I would call like a race to get out as much as possible, as quickly as possible before the war restarts,” she added.

Brohard’s claims remain unverified, although she is not the only analyst raising the possibility of undisclosed payments. “Everything appears to be happening under the radar in the Middle East, from the US convoying ships to Iran quietly charging tolls,” Chris Beauchamp, chief market analyst at IG Group, told Al Jazeera.

The possibility of secret payments would carry serious consequences if confirmed. The Trump administration has warned Tehran against imposing tolls, while Saudi Arabia, the UAE, Bahrain, Qatar and Kuwait have opposed Iranian demands for transit charges. (…)

Last month, the U.S. Treasury Department’s Office of Foreign Assets Control (OFAC) sanctioned Iranian cryptocurrency exchange BitBank, saying the Hormuz Safe Marine Services Authority had used the exchange since June to transfer payments it received to the Iranian regime. Treasury said BitBank is controlled by Iranian financier Babak Zanjani, whose network used the exchange to transfer hundreds of millions of dollars in Bitcoin to the Islamic Revolutionary Guard Corps (IRGC). Zanjani was sentenced to death in Iran in 2016 for embezzling state oil funds, but his sentence was commuted in 2024 and he later re-emerged as a backer of regime-linked economic projects. (…)

The Mystery of the Unknown Sailor Who Inspired Trump to Shake Up the Navy A chance encounter pushed the president to revamp the carrier fleet—and sent Pentagon officials on a hunt for the elusive ‘Bill Jones’

Also in today’s WSJ. You can’t make up stories like this.

(…) Jones convinced Trump the old technology was superior to modern methods of launching jets, the president said.

The incident sent Navy leaders searching for Jones to figure out what he said that prompted the president to demand an overhaul of one of the most expensive weapons programs in history. They never found him, according to defense officials. (…)

More problematic: “Steam catapults have been out of production for decades—no one makes them anymore,” said Hunter Stires, the maritime strategist to former Navy Secretary Carlos Del Toro. (…)

Shipbuilding experts worry putting steam back on a carrier would put the Navy technologically behind China.

Acting Navy Secretary Hung Cao has warned officials not to question the president’s decision. (…)

“This could ultimately be a multibillion-dollar change that was resulting from largely this one sailor’s interaction with the president,” said Bryan Clark, a naval expert with the Hudson Institute, a Washington think tank. (…)

Trump has gotten personally involved in several other projects, from the new Air Force One and a class of high-tech battleships to a futuristic military drone port on the top of the White House. (…)

Not to mention the Reflective Pool!

The rising risk of the AI-versus-everything-else economy

More and more, it seems as if there’s the AI economy — and everything else.

The growing chasm between the breakneck growth of the AI world and the slower pace of the economy where most Americans work and live has huge implications, especially for the markets.

The most recent numbers on U.S. construction spending through August — shown above — paint the picture remarkably clearly.

Data: U.S. Bureau of Labor Statistics, FactSet; Chart: Axios/Matt Phillips; Note: Data center construction costs include labor, materials, profits of contractors, architectural and engineering work, miscellaneous overhead, interest and taxes. Does not include servers, racks, chips or memory.

According to U.S. Census Bureau data, private companies are building data centers at an annualized pace of $85 billion in 2026, representing a 73% increase from 2025. Apollo Global Management notes that AI investment is accelerating at 0.85 percentage points of GDP annually.

Apollo’s Torsten Slok said consensus forecasts imply hyperscaler capital expenditures will reach about 3% of U.S. GDP annually between 2027 and 2029, up from 0.3% in 2019 and 1.4% in 2025.

ING’s calculations reveal the sluggishness of private investments other than AI:

• If we group computing and peripheral, together with software investment, we see a cumulative 62% increase in real spending since the first quarter of 2022. Other non-residential business investment fell in year-on-year terms for six consecutive quarters between 4Q 2024 and 1Q 2026.

• ING’s best estimate is that AI is responsible for 44% of US GDP growth in 2026, reduced to 36% after taking into account AI-related imports.
• ING then considers the wealth effect on consumer expenditures: “$33tn of extra household equity wealth translates into roughly $500bn of cumulative extra spending since fourth quarter 2022. Simplistically, that is around an extra $36bn of consumer spending per quarter ($144bn annualised), equivalent to around 0.65pp of 2Q 2026 consumer spending and 0.44pp of GDP which was 2.2% in Q2 and is estimated at 3.0% in Q3. The AI wealth effect could be contributing 15% of GDP growth.

All these numbers are rough estimates but confirm that the AI boom is responsible for 40-50% of US growth this year. A non-AI US economy would currently be expanding at a sluggish pace of roughly 1.0% to 1.3% annualized.

Other indications that the US economy is largely being supported by AI investments:

• Employment for people who build the equipment and parts that data centers run on — including gas turbines to generate power, wiring, electrical switchboards and other specialized components — is one of just two categories of manufacturing jobs that have consistently grown in the past few years, according to an analysis by Jiaxin He, a research assistant at the bipartisan think tank Economic Innovation Group. (Washington Post)

The One Big Beautiful Bill Act (OBBBA) also provided a significant fiscal boost to short-term economic momentum. The official CBO Dynamic Estimate calculated that the OBBBA contribution to real GDP growth would peak at 0.9 percentage points in 2026. “Because the effects of changes in aggregate demand would subside quickly, the temporary boost from demand-side factors would diminish after 2026.”

Axios continues:

The same divide is an important theme in the markets as well. In corporate bonds, AI-related offerings accounted for more than half of net investment grade issuance for the year through the end of August, according to data from Apollo and Bloomberg.

Data: Apollo, Bloomberg; Chart: Axios/Matt Phillips

The weighting of the S&P 500 technology sector in terms of market capitalization is now roughly 40%. That’s higher than during the peak of the dot-com bubble.

• But that actually underplays the impact of AI.
• If you throw in the hyperscalers Meta and Amazon, which S&P doesn’t include in the tech sector, the share of the AI-pilled part of the blue-chip index climbs closer to 50%.

Taking the cue from Kevin Warsh and listening to what the market is saying (since the August highs):

• S&P 600 Small Caps: –8.6%
• S&P Banks: –11.0%
• BAC: –17.0%
• JPM: –9.0%
• GS: –23%
• KKR: –22%
• KBW Regional Banks: –12.8%
• S&P 500 Industrials: –10.0%
• Restaurants: –14.0%
• MCD: –15.2% (-38% since March)
• Retailers: –9.6%
• HD: –18.5%

Ed Yardeni calculates that 42% of S&P 500 stocks are down year-to-date. He also shows that only 33% of stocks have outperformed YoY. The bar has been set high, however, with the S&P 500 up 12% and 14% YoY in 2025 and YtD in 2026 respectively. We have been around 30% of outperformers several times in the last 2 years. Hopefully, three’s a charm.

Looking forward, some economists argue that the rise in energy costs and short term interest rates will combine for a $600B hit to the American consumer, 2-3% of disposable income which was +4.6% in Q2’26.

Then the midterms could pit Congress vs the White House for 2 years. Impeachment? Government funding? Tax increases?

Will the war in Iran become the Forever War?

What about the whole Middle East boiling towards exploding?

Ned Davis monitors trends in supply vs demand of equities. We are at 50-50. I added the dotted verticals to NDR’s chart. Not definitive but a solid warning to be monitored.

The 13-34 week moving averages are still long-term bullish however.

(stockcharts.com)

My friend Hubert Marleau:

Meanwhile, chances of a 25-bps hike in the Fed’s policy rate at the October 27-28 FOMC meeting are only 22%, with odds of two hikes by December at just 3%, in part because the economy, in affordability terms, might not be in the best of shape to withstand any interest rate increases.

Efforts to tame the volatility of the 10-year Treasury yield, the most important number in global finance, have failed for two reasons:

1) A dramatic change in the composition of bond buyers from insensitive foreign reserve managers to sensitive private investors like hedge funds and carry-trade international players, has occurred, stemming from the weaponizing of the US dollar and out-of-control fiscal spending.

2) A huge amount of international money has flowed into the AI trade, virtually making it the new US capital market. The ratio of US net international investment position (NIIP) to the sum of all rest-of-world net creditor nations’ NIIP has surged to 80%- a clear gauge of the concentration and allotment of the world’s savings that has been sucked into dollar investments.

This has impacted the momentum of the bond market beyond what conventional theory would normally have figured, and it’s reflected in the bullish performance of the US dollar, which is perching at 102-ish, its highest level since last spring, even though the near-term hike narrative has become less threatening. (…)

Despite October’s gruesome fame, it’s actually one of the best months for stock returns. So far, the S&P 500 is green, but not immune to the rising cost of money. The economy has slowed down, but still remains in expansion territory as reflected by the September ISM index, which rose to 54.9 with higher prices received for goods and services, guaranteeing a 3% growth factor for Q3 and beyond and a positive outlook for corporate profits. (Incidentally, the latest third quarter GDP growth estimate, generated by the Atlanta Fed, is 3.6% and 2.8% for Q4.)

Without much else, a handful of AI mega-cap companies, firing on all cylinders, drove the S&P 500 to a new record high of 7,819 on Tuesday. Suddenly, all hell broke loose on Wednesday and Thursday. Oil prices surged as war was about to resume in Iran, the cost of money rose in response to the FOMC’s view that the current monetary policy is not restrictive, there was talk of a potential French debt crisis not being far off for Les Misérables, and the rapid weakening of the yen, combined with rising Japanese interest rates, accelerated the unwinding of the yen-carry trade, sparking fresh selling in both bonds and stocks.

Nonetheless, the S&P 500 held up remarkably well for the week as a whole, up 1.0% to close at 7811, despite this collection of forces, which is difficult to dismiss as mere background noise.

What is most important is that the AI boom remains powerful, accelerating the expected year-over-year increase in S&P 500 earnings per share for Q3 to 27%, after registering blockbuster gains of 30% and 25% in Q1 and Q2, respectively, while fabulous earnings are broadening across corporate America.

Analysts expect all 11 S&P 500 sectors to deliver positive y/y growth in both revenues (11.6%) and earnings (30.6%). And this phenomenon is not about to stop, according to Yardeni Research, which came out this week with its earnings-per-share prospects for the S&P 500 over the remainder of the decade, projecting $450 per share in 2028, $475 in 2029, and $500 in 2030, which supports a S&P 500 target of 10,000 before the end of 2029.

It should be added that the street is more bullish on earnings per share, forecasting $419 in 2027 and $489 in 2028. These elevated long-term growth assumptions are enough for me to keep my year-end target of 7,925 and 8,700 for 2027.

• Goldman Strategists Upgrade Global Stocks on Earnings Boost

Equity prices are less demanding after higher oil prices and bond yields stalled the rally, while the outlook for profits is still healthy, the team led by Christian Mueller-Glissmann wrote in a note. They raised their three-month view on stocks to overweight from neutral, matching their bullish stance over a 12-month timeline.

“Earnings growth remains supportive for equities and the recent valuation de-rating has improved asymmetry into the third-quarter earnings season,” Mueller-Glissmann said. (…)

Analysts expect S&P 500 companies to post a 25% jump in third-quarter profits compared with a year earlier, among the best rates of increase on record, according to data compiled by Bloomberg Intelligence.

Other Wall Street strategists, including at Bank of America Corp and Barclays Plc, have also turned more bullish about the outlook for US stocks recently, citing among other things a robust earnings outlook. (…)

FYI:

Q3’26 earnings are another K-shaped trend per LSEG IBES: 5 sectors, all either commodity (war) or AI-related are expected to post earnings growth or 57.4%. The other 6 sectors, (consumer discretionary and staples, financials, health care, real estate and utilities) are seen growing earnings 4.2% on average.

For all of 2027, total S&P 500 earnings are indeed forecast to rise 15.9%. Only 3 sectors will beat that however, growing earnings 26.8% on average. The other 8 sectors are expected to post earnings up only 2.4% on average.

That’s in an economy still strongly supported by AI:

The Desperate Hunt for AI Computing Power Is Upending Silicon Valley

The WSJ paints the hunt for scarce compute power. The main points:

• Compute, the industry shorthand for computing power [i.e. GPUs, CPUs], is the lifeblood of the AI boom, and it’s in short supply. Now fights over how to get more of it—and what projects to give priority to—are sending a powerful industry scrambling. (…) “There’s a big gold rush happening at the moment. It’s hard to fully comprehend the scale.”
• Established suppliers of computing power are requiring customers to sign multiyear contracts with heftier down payments. Startups are increasingly inking deals with unreliable providers out of desperation, like ones lacking proper security or weaker infrastructure.
• With so much demand, prices are going up. The hourly cost of renting computing power from Nvidia H100 chips [2022 vintage] on a one-year contract has risen 60% in the past year, according to SemiAnalysis, a data and consulting firm.
• To hedge against the unpredictable nature of securing computing power, some companies are signing contracts for more capacity than they need. In other cases, startups have predicted surging demand for their AI products that never materialized, forcing them to resell leftover capacity. This confluence of factors coupled with rising wait times and skyrocketing costs of compute are crippling a generation of newcomers trying to compete with Google, Anthropic and OpenAI.
• Some venture-capital firms are now reserving clusters themselves and renting the capacity back to their startups, or offering to guarantee contracts with compute providers.
• “There’s a lot of capital floating around for these companies, but there’s not a lot of deep intelligence on how to build these things.”

The need for compute has exploded since “Claude” launch last spring. AI agents are quickly becoming ubiquitous. Wait till we get autonomous cars and robots constantly monitoring their environment and talking among themselves. It’s actually very, very hard to fully comprehend the scale.

Nathaniel Whittemore’s Oct. 6 AI Daily Brief episode discussed whether “Consumers will never pay for AI” (starts at 13 min.).

In 1992, I was explaining to an audience of smart, busy investors why my equity fund included fledging cellular-related companies. When I said that in the not-too-distant future everybody would have a small phone tucked to their belt or inside their purse and that we could all be reached 24/7, the crowd exploded: “No way!”, “Over my dead body”!

It took 6.5 years for smartphone penetration to grow from 2% to 50%. It’s now 91% of all US adults.

ChatGPT went from 900 million weekly active users in early 2026 to 1.2B now. There were 35M paid subscribers in July 2025. That jumped 43% to 50M in February 2026 and OpenAi targets 122M by the end of 2026 when AI apps will only get started.

My son Danny has been experimenting with the recently launched ChatGPT DOT personal agent. Often quick to find flaws in new apps, yesterday he wrote: “A real game changer. The biggest evolution since the arrival of ChatGPT.”

According to The Information, only about 4.5% of eligible U.S. consumers currently pay for a premium AI subscription (across ChatGPT, Claude, or Gemini).

Most are using it for simple queries. Soon, they will start using agents for diverse tasks.

I recently queried Gemini to give me trends on Chinese export data for a few specific commodities. It gave me the data and built the charts. Then, to my surprise, it casually offered to create an agent for me that would monitor these trends and alert me if some pre-specified events occur. That agent would be constantly working in the background, monitoring the data for me, and pinging me on certain events.

Imagine…

Examples of surging, unmet demand keep coming:

Nvidia-Backed Lumentum Sold Out of AI Server Parts Till 2029

Lumentum Holdings Inc.’s optical components are “completely sold out” through early 2029 on demand from tech companies clamoring for faster AI data centers, according to its chief executive officer.

Lumentum is unable to meet about 70% of demand for some products through next year, and won’t be able to address 30% of demand for some other products through 2028, Michael Hurlston said in an interview in Tokyo Friday. Customers are already starting to lock in orders for 2030, he added.

The company, which received an investment of $2 billion from Nvidia Corp. earlier this year along with rival Coherent Corp., supplies advanced indium phosphide devices that help enable high-speed cloud computing and data transmission. Analysts on average expect Lumentum’s sales to more than double this year.

“We are really trying to catch up, add capacity as fast as we can, but we are very, very far behind what customers want,” Hurlston told Bloomberg News. “We are completely sold out, so there’s no end in sight.” (…)

Hurlston shrugged off fears of a pullback. The buildout remains intact, and the sector is on the cusp of a bigger opportunity when data centers adopt optical technologies at shorter distances to AI accelerators, he said.

Silicon photonics, which uses light to transmit data, is expected to replace copper-based interconnects in data center clusters as advanced AI hardware demands faster transmission. Optical connections, which generate less heat and consume less power, are also moving progressively closer to the chips. In addition to connecting data-center racks to one another, they are increasingly eyed as ways to link tiny components on the same circuit board.

Lumentum offers a type of laser that’s used to encode data onto light near the chip. Since that means high-speed electrical signals only have to travel a few millimeters, that cuts power loss and heat emission even further.

“This market can be our biggest by far,” Hurlston said, adding that demand for lasers in next-generation AI systems is expected to gain traction around 2029. “It’s another order of magnitude in terms of size.”

The sector is also bracing for further supply constraints as the US reportedly considers a ban on new Chinese optical transceiver models. Lumentum and Coherent would not be able to fill the vacuum left by Chinese makers such as Innolight, according to Counterpoint Research. Roughly two-thirds of the world’s high-speed AI optical transceivers come from Chinese makers. China also remains a big source of indium-phosphide substrates used to make lasers and other optical components.

The US needs to balance national security concerns with the risk of disrupting the supply chain, Hurlston said.

To meet demand, the San Jose, California-based company has ramped up capacity 12-fold at its Sagamihara facility — one of the world’s most advanced indium-phosphide device production sites — in the Tokyo metropolitan area over the past two years. It plans to invest at least $350 million in the site and a neighboring facility, of which it’s deployed $75 million so far. It’s also expanded capacity at its plant in Caswell, England. (…)

It takes three to five years to build new production capacity (…).

One factor that allays Lumentum’s concerns about fast-paced ramp ups is the willingness of hyperscalers to take on some of the risk that suppliers take on when they build new production facilities.

“These hyperscalers are willing to put money behind it,” Hurlston said.

In case you wonder, Lumentum is also expanding in the US:

• In March 2026, Lumentum announced the acquisition and multi-hundred-million-dollar retrofit of a 240,000-square-foot production facility in Greensboro, NC. Acquired from Qorvo, this fab is dedicated to manufacturing advanced indium phosphide (InP)-based lasers and optical devices. It is expected to fully ramp up production in mid-2028.
• Lumentum maintains an active semiconductor facility at its headquarters in San Jose, CA. In August 2025, the company funded a massive capacity expansion here to produce ultra-high-power (UHP) lasers used in next-generation Co-Packaged Optics (CPO) platforms.
• A $2 billion investment and purchasing agreement with NVIDIA has financially anchored Lumentum’s U.S. fabrication roadmap. NVIDIA is a primary customer backing these U.S. facilities to ensure it has domestic access to critical optical components.

How the AI Boom Ends Demand isn’t about to let up, but prices will start plunging by factors of 10.

Andy Kessler in today’s WSJ:

I can tell you how the artificial-intelligence data-center boom will probably end, but I can’t tell you when. (…)

AI token prices have been dropping but are set to plunge faster than most anticipate. (…)

Last week, according to OpenRouter, OpenAI and Anthropic were charging a premium for some of their models—$1 to $2 per million input tokens. Comparable Chinese open-weight models like Z.ai’s GLM and Alibaba’s Qwen charged 6 to 12 cents, around 90% less—that factor of 10 inflection point. (…)

AI architecture is evolving. Instead of a trillion parameters, efficient small language models might use 70 billion and run on an Apple Mac. These local models have no token costs beyond the cost of hardware and electricity.

But they must be less capable than the large language models, right? Not really. Stanford researchers show that a locally run small language model “can successfully answer 88.7% of single-turn chat and reasoning queries.” No question we’ll always have high-end models running in data centers. IBM still sells mainframes. But growth will happen elsewhere.

Where? Another Silicon Valley saying I’ve lived by: Intelligence moves out to the edge of the network. The real breakthrough will happen when robust AI models run on smartphones. It’s already starting. In September, California-based PrismML released a compressed version of the popular open-source Qwen3.8 27B AI model. The company got it down to under 6 gigabytes. That means it fits not only on PCs but also on high-end smartphones. The memory size is reduced by roughly 9 to 10 times. There’s that factor of 10.

Technology doesn’t stand still. Qualcomm recently announced a new Snapdragon neural processing unit, or NPU, for smartphones. Using what the company calls a “mixture-of-experts” model, it could run a 30 billion parameter AI model in 3 gigabytes of memory. According to the Deep View, the largest AI model on a phone has been four billion parameters on an Apple A20 Pro chip. Qualcomm provides a 7.5 times boost.

NPUs, like Qualcomm’s new Snapdragon and Google’s Tensor chips in its Pixel phones, are multiplying. Same for the latest Apple M- and A-series chips, engines that can run AI models locally on iPhones—maybe why Apple hasn’t joined the AI data-center spending delirium.

A Nvidia-backed startup, Reflection AI, claims it matches leading Chinese open-weight AI models with “3-4x less inference compute.” That’s it? Only 3 or 4 times? Wake me up when it’s 10 times. At the pace AI is moving, that could be year-end.

Agents are all the rage, from Meta’s Muse to xAI’s Grokbot. They run on virtual machines in data centers but raise all sorts of privacy concerns. Startup Perplexity recently teamed up with Nvidia to run agents on PCs. Zero token costs. And you own your data.

I often hear, that if tokens get cheaper, demand will go up. That’s Jevons Paradox. Sure, elasticity works. But if prices drop 90%, demand must increase 900% just to stay flat. Remember when internet demand grew while debt-laden Global Crossing spectacularly failed because prices dropped too fast? Historically, semiconductors got 30% cheaper every year. When prices dropped by more than 50%, disaster struck with fabrication plant closings and company failures.

Will 2027 see peak data center? Again, timing is dicey. Current demand is gangbusters, but spending is a game of chicken until multiple factor-of-10 improvements kick in and prices collapse. This could set off a cascade of data-center halts and debt defaults, even as token demand increases.

Yes, model training, drug discovery and winning prizes like the recently solved Navier-Stokes equations require hot chips in data centers. But chatbots and agents and other tasks may increasingly be handled on PCs and phones and smart glasses and whatever new gizmos Silicon Valley invents over the next decade.

I’ve heard this called the “Great Migration.” Mr. Jurvetson’s factor of 10 is visible now, as it often is, but ignored. Change doesn’t happen overnight. It happens gradually, then suddenly.

Had Andy used AI to criticize his own article he might have read things like this:

• While the piece correctly identifies strong historical tech patterns (like downsizing and cost deflation), its core thesis suffers from several structural flaws, economic misinterpretations, and questionable technical assumptions.
• The author attempts to invoke Jevons Paradox but completely flips its economic meaning. The author claims that if prices drop 90%, demand must increase 900% just for revenue to “stay flat,” framing this as a steep hill to climb. However, historically in digital tech, a 90% drop in computing or data costs triggers a 10,000% or greater explosion in demand. When data became cheap, we didn’t just send more text emails; we started streaming 4K video. Lower token prices don’t collapse the market; they make embedding AI into every micro-interaction economically viable.
• The article pits edge computing (on-device) against data centers as a zero-sum game, assuming edge growth will cannibalize data center demand. On-device AI and cloud data centers handle fundamentally different workloads. Running a 30-billion-parameter compressed model locally is great for basic text tasks, predictive typing, or localized voice assistants. However, it cannot replace data centers for massive multi-modal workloads, heavy synthetic data generation, real-time complex reasoning (like multi-step coding agents), or continuous foundational training. The “Great Migration” is actually a hybrid model: edge devices will act as filters for simple tasks, while data centers handle the complex, heavy-lifting cognitive tasks.
• About the Stanford 88.7% figure: The study covers only single-turn chat and excludes agent and long-document tasks, where local models trail by up to 45 points.
• Apple isn’t skipping data centers; it’s renting them. Its new Siri runs on a custom 1.2-trillion-parameter Gemini model, partly on Google Cloud Nvidia Blackwell GPUs. Apple’s restraint moves the spending to Google; it doesn’t remove it.
• Perplexity’s “zero token costs” is great for simple tasks. But it sends harder steps to the cloud with a subscription.” Owning your own data” will only encourage usage of AI agents.
• The article assumes data centers are built primarily to serve the current level of chatbots. The data center boom is primarily driven by an aggressive arms race to train next-generation frontier models and run large-scale simulation environments. Even if inference completely migrates to the edge, the computing power required to train the next generation of models continues to grow exponentially, meaning data center demand will remain robust.
• The author selectively picks historical tech cycles to fit a narrative of impending doom. While Global Crossing defaulted during the dot-com crash, the physical fiber-optic cables they laid down laid the foundation for the modern internet boom. Similarly, even if some current data center debt structures face turbulence, the physical infrastructure (gigawatts of power, fiber, and cooling) will remain highly valuable assets, not useless relics.

FYI: Google’s monthly token volume went from 480 trillion to 3.2 quadrillion in one year, a 7x jump, and 330x over two years. That was only Google, before agents, autonomous cars, robots, etc.

China Passenger Vehicle Retail Sales Drop 24% in September

China’s passenger vehicle retail sales fell 24% in September from a year earlier to 1.702 million units, while new-energy vehicle sales declined 12% to 1.141 million units, according to data released by the China Passenger Car Association.

Cumulative retail sales for 2026 reached 13.42 million units, down 21% from the same period a year earlier, the association said. New-energy vehicle retail sales increased 14% from the previous month, with year-to-date sales totaling 7.816 million units. They were down 12% on year.

The association said new-energy vehicles accounted for 67.1% of passenger vehicle retail sales in September, while their share of wholesale volumes reached 66.1%.

• China’s Geely Automobile to Enter Canadian EV Market Next Year

Geely Automobile Holdings Ltd. plans to launch in Canada next year, positioning it be the first Chinese mass market electric vehicle maker to establish a presence in the North American market.

The carmaker has begun setting up a retail and service network in the country ahead of the 2027 start, according to a company statement Friday. Geely, which owns premium EV brands including Zeekr and Lynk & Co as well as more affordable offerings, didn’t specify a start date for sales or which models would be available.

The move marks a significant step after Ottawa and Beijing reached an agreement in January to allow a limited number of Chinese EV exports to Canada. As trade disputes deepen with the US, Canadian Prime Minister Mark Carney’s administration is looking to forge new alliances, including by allowing Chinese automakers in and hoping to attract investments from them. (…)

Chinese rivals BYD Co. and Chery Automobile Co. are also wooing dealers and setting up distribution. (…)

Canada Employment Falls by 68,300 Amid Public Sector Decline

The unemployment rate ticked up to 6.5% in September from 6.4% in August, Statistics Canada reported on Friday.

The job loss reflected a decrease in public sector employment, which fell by 70,000. Education led the decline, with 35,300 fewer people working the sector. Employment also fell in healthcare and social assistance by 23,100.

Manufacturing saw a decrease in employment of 12,700.

While the job losses last month were driven by the public sector, the cumulative decline in employment since the start of the year paints a sluggish picture of the labor market. Employment fell in Canada by 41,200 since December 2025, the largest year-to-date loss since the 2009 financial crisis, excluding 2020 during the Covid-19 pandemic.

The September data also gives the first glimpse of the impact of so-called Section 338 tariffs on the Canadian economy. New 50% US tariffs on $20 billion of Canadian goods took effect on Aug. 22 after the breakdown in trade negotiations between the two countries, while retaliatory tariffs on US goods kicked in on Sept. 8. Job losses in August and September total 110,000. (…)

Friday’s report also shows the participation rate declined by 0.2 percentage points to 64.8%, marking the lowest level since December 1997 — excluding 2020 during the pandemic.

StatCan says the decline in the participation rate can be attributed largely to population aging, and says this trend is expected to continue, despite older Canadians participating in the labor market at higher rates compared to previous cohorts. (…)

The average hourly wage for full-time permanent employees rose by 2.3% from a year ago, matching economists’ expectations, and up from 2% the previous month.

The decline in employment was driven entirely by a significant drop in public-sector employment. (@RichardDias_CFA)

Deadly Attacks on Saudi Arabia Risk Wider War That Draws In US

This weekend’s deadly attacks on Riyadh’s main airport have demonstrated the ability of Yemen’s Houthis to strike at the heart of Saudi power, raising the risk of a dangerous escalation that could draw in the US. (…)

The attacks put de facto ruler Crown Prince Mohammed bin Salman under mounting pressure to escalate a war the country has spent years trying to escape, against an opponent that’s arguably more capable than it’s ever been. They also present difficult choices for the US and the kingdom’s new defense partners, Turkey and Pakistan, over how far to go in helping Saudi Arabia respond. (…)

Saudi Arabia’s Foreign Minister Prince Faisal bin Farhan and his Iranian counterpart Abbas Araghchi discussed the latest escalation in the region in a phone call, according to a report Sunday by the state-run Saudi Press Agency. (…)

The group has survived years of Saudi and US bombing campaigns, and the latest strikes demonstrate how difficult it may be to stop it from inflicting further damage on the kingdom. (…)

“It is very unlikely that the Saudis can stop the attacks if the US could not — short of sending in ground forces or dramatically improving the capability of the Yemeni fighters it is supporting,” Pettyjohn said. “Suppression requires constant air patrols searching for launchers, which is difficult. And the Yemenis seem to have used the time since the ceasefire with the US to rearm.” (…)

The escalation also poses a test for Turkey and Pakistan, which signed a defense pact with Saudi Arabia in August, known as the Mecca Defense Alliance. Both have signaled willingness to send troops or weapons to help defend the kingdom, but Turkey has ruled out joining offensive operations against the Houthis. (…)