KLA is the 800-pound gorilla of chip inspection. But tiny Onto Innovation is growing its advanced packaging business 80% this year by catching defects in AI memory stacks — and just took share with a product that runs 3X faster. Sometimes the best trade isn't the gorilla, it's the specialist.

Everyone's watching Nvidia vs AMD. Meanwhile, both of their next chips depend on one thing: getting a giant multi-chip package built without a single defect. Whoever sells the inspection tools gets paid either way. Metrology is the quietest AI trade out there. 🔬

What good is a faster AI chip if the network can’t keep up? $CLS has 10 active 1.6T networking programs, with growth expected to accelerate in 2027. AI’s buildout reaches far beyond GPUs. The connections between the chips deserve attention.

For 40 years, stocks and bond yields moved opposite each other. That's what the 60/40 portfolio was built on. From the 1940s to the '70s, they moved together, and that's happening again. With this much debt, every path leads to more borrowing and more money printing. Hard assets and great businesses can't be inflated away. Fixed-income bonds can

The S&P is still in an uptrend while SPY holds 761, and I think there's one more high in it. After that I expect a healthy pullback into early November, maybe to 750–740, and I'll be buying it. The line I'm watching is SMH at 545. If semis lose it, it won't be a quiet move. 📉➡️📈

An AI chatbot gives you an answer. An AI agent can change your database or execute code. That changes the security problem completely. $NVDA wants to supply the controls around those digital workers. AI adoption could create demand far beyond chips.

This bull market is 17 years old and up 1,000%+. The average secular bull lasts 11 years and gains 675%. Everyone says that means it's over. History says the outliers run longer than anyone expects. S&P 9,000–10,000 isn't crazy. 🐂

The AI build-out isn't just GPUs. It's the pipes. $CLS has ~41% of the high-speed Ethernet switch market, and 1.6T is about to surge. When data centers upgrade networking, this is who gets the call. 📈

Memory used to be a guessing game: build capacity, hope demand arrives. Long-term customer deals could change that for $MU. The catch? Price ceilings can limit upside when supply gets tight. Would you trade some upside for years of better visibility?

In the cloud era, CrowdStrike became the security guard for your laptop. In the agent era, it's trying to be the security guard for your AI. Estimates put it at ~90 agents per employee. Each one needs to be watched, secured, and budgeted. That's a whole new market, and $CRWD is billing for it separately.

The next AI battle: how cheaply can companies RUN it? HPE says its private-cloud AI can cut token costs by up to 60% vs public cloud, based on internal analysis. If customers see those savings, servers + networking + storage get interesting. $HPE’s pitch: cheaper AI.

Unpopular take on $MU: the 343% revenue growth quarter is gone — and that's bullish, not bearish. Micron traded explosive-but-unstable growth for 5-year contracts (SCAs) that lock in pricing floors AND still let next-gen HBM/DDR6 chips price at market. Slower headline growth, steadier cash flow, longer runway. That's not a top — that's maturity.

Dell FQ2, in one breath: • AI server backlog: $95B (+8x YoY) • FY guidance raised $25B... again ($52B raised in 2 quarters) • EPS up 273% YoY • Operating margin: best in 10+ years And AI servers weren't even half of the raise. This is what "hitting on all cylinders" looks like.

$MPWR may be turning into one of AI's ultimate picks-and-shovels plays. It can benefit from: → Nvidia Rubin → AMD next-gen accelerators → x86 CPUs → Arm CPUs → 800G → 1.6T optics → NICs + DPUs → Vertical power delivery → Future 800V data centers Enterprise Data: +164% YoY Communications: +78% YoY Total revenue: +48% YoY The AI buildout is expanding far beyond GPUs.

Small company. Very big inflection. $AEHR: • ~$50M trailing revenue • $130M-$150M FY27 guide • Up to 200% growth • Q4 bookings +447% YoY • Backlog 5X+ YoY • ~95% of revenue now AI • Major AI chip + hyperscaler customers ramping The question is no longer whether AEHR escaped the EV slowdown. It’s whether AI just created an entirely bigger company.

The real Nvidia story isn't chips anymore. It's power. Revenue per gigawatt: $18B (Hopper) → $25B (Blackwell) → $40B (Vera Rubin). Same power footprint, 2.2x the revenue. Early tests show Vera Rubin pushing 10x more tokens per megawatt vs the current gen. Whoever wins the power game wins AI.

The most interesting ticker in the OpenAI Jalapeño story might not be $NVDA. It might be $AVGO. OpenAI + Broadcom just demonstrated that custom AI silicon can potentially compete with the world's best merchant GPUs on the metric hyperscalers increasingly care about: Tokens per MW. Training may remain GPU dominated. But inference is becoming enormous—and power constrained. If hyperscalers increasingly conclude: “Why pay Nvidia's margins when we can design silicon around our own workloads?” Broadcom's custom ASIC opportunity could be much larger than the market imagined. Jalapeño isn't the death of Nvidia. But it may be validation of the custom-silicon era.

10GW of AI datacenters is not a GPU story. It's a: • GPU story • HBM story • networking story • optics story • power-generation story • transformer story • cooling story • natural-gas story SemiAnalysis believes SpaceX could build ~10GW of AI compute by YE2027 and potentially monetize just half of the incremental capacity into roughly $300B ARR. Whether that exact number happens almost misses the point. AI is colliding with the physical limits of the real world. The next winners may be whoever removes those bottlenecks.

Google may have made one of the strangest—and potentially most profitable—decisions in AI. Instead of hoarding compute to make Gemini #1… it's selling that compute to Gemini's competitors. SemiAnalysis estimates 20%+ of TPU shipments from Q3'26–Q4'27 go directly to Anthropic. Bad strategy for winning AGI? Maybe. Great strategy for monetizing the AI boom? Possibly spectacular. $GOOGL might lose the model war and still win the economics.

A year ago, "open-source AI" meant "cheaper but behind." Not anymore. Kimi K3 (out of China) is now the top open frontier model — and it got there by solving a problem the giants hadn't cracked yet: how to keep AI fast and cheap during hours-long agentic sessions, not just quick Q&A. The frontier isn't closed anymore.