Editor's Note

The period started with disputed rate-hike odds and ended with stronger jobs, high services prices, and expensive energy keeping a September hike in play.

AI demand did not slow down. But the bottlenecks became clearer: power, memory, data centers, cooling, and customer concentration.

The simple investor question is no longer just whether AI grows. It is whether the growth turns into durable revenue and cash flow fast enough.

Rates Became A Bigger Problem Again

The rate story hardened through the period.

Early on, Kalshi put the chance of a rate hike next month at 53%, while Watcher.Guru reported 49%. That was disputed. Gurgavin said there was "NO WAY WE SEE A HIKE NEXT MONTH."

By later in the period, markets were pricing roughly a two-thirds chance of a September hike. Then jobs came in stronger than expected.

The numbers that mattered:

• U.S. added 162,000 jobs in August.
• Unemployment held at 4.1%.
• ISM Services PMI rose to 55.4.
• ISM services prices-paid reached 72.6.
• WTI closed above $90.
• Brent traded near $95.
• Retail diesel reached $5.78 per gallon.

That mix matters because it gives the Fed less reason to rush toward easier policy.

Fed Governor Christopher Waller said he supported holding rates at 3.50% to 3.75% if inflation kept moving toward the Fed's 2% target. He also said an inflation reversal could make an increase appropriate.

The next tests are close: CPI and PPI on Sept 10 and Sept 11, then the Sept 15 to Sept 16 Fed meeting.

The surprise was not that rates were discussed. The surprise was how quickly oil, services prices, and a strong jobs report made rate hikes feel live again.

AI Growth Ran Into Physical Limits

AI demand kept growing, but the week made one thing clearer: chips are not enough.

The reported consensus estimate is that about 15GW of AI compute produced in 2027 cannot be turned on in 2027. Deployment also requires transformers, wiring, liquid cooling, massive chillers, and complex networking.

Power became part of the AI race.

President Trump declared a national emergency and signed Executive Order 14420, banning foreign-made equipment from the U.S. power grid. The Department of Energy has 120 days to decide which foreign transformers, inverters, and control systems will be removed.

Anthropic reportedly beat Google and Microsoft for Nscale's West Virginia campus in a $45 billion deal. The supplied material described the site as evidence that bidding has moved from chips to interconnection queues.

That is a useful phrase because it explains what changed. Existing grid capacity can matter as much as hardware.

Vistra was one public example of that theme:

• Approximately 44,000 MW of generation capacity.
• 20-year power agreements with Meta and Amazon totaling 3,800 MW.
• Q1 2026 revenue of $5.64B, up 43.4% year over year.
• Net income of $1.029B, compared with a loss in 2025.
• Planned acquisition of a 5,500 MW Cogentrix Energy natural-gas portfolio for approximately $4.7B.

The reported long-term need is large: the U.S. requires between 50 and 100 GW of new electricity capacity by 2030 for AI data centers alone.

AI infrastructure is becoming an electricity story, not just a semiconductor story.

Broadcom Put Real Numbers On AI Demand

Broadcom gave investors the clearest public numbers on AI semiconductor demand.

Q3 FY26 results:

• Revenue: $29.59B versus $29.43B estimated.
• EPS: $3.32 versus $3.24 estimated.
• Free cash flow: $13.7B versus $13.8B estimated.
• Adjusted EBITDA: 66% of revenue.

The AI number was much bigger:

• AI semiconductor revenue reached $16.7B, up 221% year over year.
• Q4 AI semiconductor guidance was $21.7B, up 236% year over year.
• Broadcom targets roughly $58B in FY26, $115B in FY27, and $230B in FY28 AI semiconductor revenue.
• The company expressed a "high degree of confidence" that it will ship $350B of AI semiconductors over the next 2 years.

Shares still fell 6.6%.

That reaction is important. The market did not simply reward the big AI numbers. Investors also focused on the near-term guide and how much future growth depends on large AI customers.

Cantor Fitzgerald raised its Broadcom price target to $600, based on 17 times its CY28 EPS estimate of $36. The note expects about 20GW of FY28 customer demand, including approximately 10GW from Anthropic, approximately 5GW from OpenAI, approximately 5GW from Google, and more than 1GW from Meta.

The open issue is concentration. Kalshi reported that 80% of OpenAI and Anthropic's enterprise revenue comes from 1% of customers.

Broadcom's numbers are real. The question is how much of the AI buildout depends on a small number of buyers continuing to spend at very large scale.

Memory Became The Other Bottleneck

The memory shortage did not ease in the supplied material.

Reported signals were severe:

• SK Hynix says it is sold out past 2030.
• Micron has customers locked through the end of the decade.
• Apple called the shortage a "hundred-year flood" and raised iPad and Mac prices.
• NVIDIA increased server prices about 15%.
• NVIDIA lifted memory-heavy commitments from $119 billion to $279 billion.

UBS projected Micron revenue could reach nearly $379 billion in fiscal 2028, with profits approaching $286 billion and EPS of approximately $265.65. That compares with projected fiscal 2026 EPS of $74.13.

The bull case rests on tight supply. Micron says its 2026 HBM supply is fully committed, and UBS estimates the HBM market could reach approximately $100 billion by 2028.

The pattern is simple: AI demand is pulling on memory, not just GPUs.

Software, Security, And Robotics Joined The AI Spending Story

The AI story was not only chips and power.

Snowflake reported 37% product revenue growth, raised guidance, and said AI produced roughly half of its recent growth acceleration.

ServiceNow signed a collaboration agreement with Aramco Digital to power AI workflows across Aramco's affiliate network in 50+ countries. HPE expanded its Oracle AI data center agreement and reported Q3'26 revenue of $12.2B, up 34% year over year.

Cybersecurity also moved deeper into AI. CrowdStrike launched SafeMind, a new family of cybersecurity AI models built with NVIDIA's Nemotron and powered by CoreWeave. Palo Alto's Nikesh Arora said AI is going to be weaponized by bad actors, while cybersecurity infrastructure deployed 7-10 years ago is not built to handle machine-speed attacks.

Robotics moved from talk to early deployment.

Tesla started offering Cybercab rides to the public in limited areas of Austin, Texas. The Cybercab has no steering wheel or pedals. Tesla registered 45 Cybercabs in Texas, bringing its state robotaxi fleet to 420 vehicles.

Figure committed $3.5B to deploy up to 100,000 Nvidia Vera Rubin GPUs for humanoid robotics and Helix AI. Nvidia CEO Jensen Huang said physical AI could be "10x larger than digital AI."

These are different markets, but the pattern is the same: more AI use cases mean more demand for compute, memory, power, and software.

Paper Gains Made The AI Profit Picture Messier

One number should make investors slow down: Big Tech companies booked more than $160bn in gains last quarter from investments in other AI companies.

Those gains lifted reported earnings. They also raised the question of how much of the AI boom is operating profit and how much is paper profit.

Alibaba showed another version of the same tension. It reported 9% growth, with cloud up 45% and AI up 16%. But operating activities provided 3 billion, capital expenditures were 9 billion, and cash declined from 60 billion to 30 billion.

The speaker behind the valuation review was not willing to pay for those promises. He concluded: reprice Alibaba, look for better situations, and do not bet on AI.

That is the investor test across the whole period. AI revenue is showing up in places like Broadcom and Snowflake. But AI spending, paper gains, and high expectations are showing up too.

Counter-Thesis and Risk Watch

LongGameEquity called cybersecurity a strong AI opportunity but said valuations are becoming "absolutely ridiculous." The cited forward P/E and price-to-sales figures were:

• CrowdStrike: 186x forward P/E and 40x price-to-sales.
• Cloudflare: 209x and 43x.
• Palo Alto Networks: 87x and 21x.
• Rubrik: 327x and 11.5x.
• Zscaler: 38x and 8x.

LongGameEquity's conclusion was: "Great companies ≠ great stocks at any price."

Investing Visuals separately called CrowdStrike an amazing company but cited 110x next-12-month EV/EBITDA for 26% revenue growth. The author said nearly all the upside felt priced in.

Palo Alto's Nikesh Arora warned about neoclouds. He said, "In 2 years from now you will be able to buy a neocloud for less than they raise at today." Shay Boloor said the warning could apply especially to businesses with land, power, or buildings but without the infrastructure and operating scale to deploy computing efficiently.

A risk-watch post noted that SPY was about 1% below its all-time high while several high-beta growth stocks had already fallen sharply, including IREN down 53%, ASTS down 56%, RKLB down 57%, and TE down 60%. The source asked what would happen to those stocks if SPY fell 10% to 20%.

The Wall Street Millennial channel argued that Chinese memory supply could eventually reverse the Micron bull case. The speaker said DRAM prices increased 500% over the past 12 months, and expects a historic memory glut could arrive within the next 1 to two years.

ClearValueTax framed the strong jobs report as bad news for stocks because it strengthened the case for higher rates.

Looking Ahead

The next rate tests are already set: CPI on Sept 10, PPI on Sept 11, and the Fed meeting on Sept 15 to Sept 16.

For AI, the open questions are more business-specific:

Can Broadcom's large AI targets turn into revenue without disappointment around customer concentration?

Does the memory shortage stay tight enough to support the Micron bull case?

Can AI software revenue grow fast enough to offset the spending, stock compensation, and paper-profit questions now showing up across the sector?