Blue | Semis & AI Infra

B.S. + M.S. in CS | Primary-source research on AI infrastructure, semis & memory cycles. Finding the bottlenecks before they hit consensus.

[Blue | Semis & AI Infra @BlueTradeIn]

The AI bull case deserves more scrutiny on red days. It doesn’t deserve to be rewritten by every red candle.
Higher bond yields are a real problem. They raise financing costs and reduce what investors will pay today for tomorrow’s profits. AI stocks can fall even while the businesses keep growing.
But look at what customers are doing.
On September 22, Nvidia announced that Sea is adopting Vera Rubin to expand AI across its businesses. Sea already uses AI in shopping, fraud detection and gaming. That is a concrete example of adoption moving beyond model developers into everyday services. Source
The longer-term opportunity is straightforward: when AI improves a customer’s economics, that customer has a reason to deploy more of it. If usage grows faster than efficiency reduces compute per task, total infrastructure demand expands.
That is the case worth tracking for $NVDA , $AMD and $MU . Each still has to turn demand into profitable sales.
Watch orders, utilization, margins and cash flow. If customers start cancelling capacity or returns disappoint, reassess. A lower share price alone tells you none of that.
There is still good reason to be optimistic about AI infrastructure. Just leave room for valuation risk and a rough ride.
You don’t need to believe every dip is the bottom to believe the industry has years of growth ahead.

23:41

[Blue | Semis & AI Infra @BlueTradeIn]

A fixed power budget makes efficient networking more valuable.
$MRVL ’s September 21 announcement includes 2nm optical technology demonstrations targeting 400G per lane and the transition toward 3.2T connectivity.
The investment logic goes beyond faster connections.
AI clusters spend power moving data as well as processing it. As bandwidth requirements rise, keeping energy per bit unchanged makes the networking bill harder to accommodate.
Reducing that cost can free up power for useful computation within the same facility limit.
That gives customers a reason to upgrade connectivity even when adding more electricity is difficult.
These are demonstrations, not proof of volume revenue. Customer qualification, reliability and production economics still have to follow.
But Marvell is targeting a problem that gets more expensive as AI infrastructure grows: moving more data without letting the network consume an ever-larger share of the power budget.

[Blue | Semis & AI Infra @BlueTradeIn]

Nvidia is giving robot builders free software. There’s a hardware business behind that generosity.
On September 22, $NVDA released Isaac ROS 5.0, adding AI-agent development tools and support across Jetson platforms, from Orin Nano to Thor.
The investment logic is straightforward: make robots easier to build, then compete to supply the compute inside them.
A working robot has to process camera feeds, track objects and plan movements within tight response times. Those workloads create demand for onboard computing, beyond the data centers used to develop its models.
If better tools help more projects reach commercial deployment, Nvidia gains another route to hardware sales.
That’s the opportunity: AI compute spreading from server racks into machines that do physical work.
A software release doesn’t guarantee robot orders. But lowering the engineering barrier gives more builders a reason to start on Nvidia’s platform.
The tools are free. Deploying a fleet isn’t.

[Blue | Semis & AI Infra @BlueTradeIn]

$AMD crossed $1 trillion on September 21. The useful question is what customers will buy next.
AMD has two ways into the AI budget: Instinct GPUs for model computation and EPYC CPUs for the servers supporting it.
An agent generates code, then somebody’s infrastructure has to execute it, run tests, retrieve data and manage the workflow. Much of that work runs on CPUs.
That gives AMD an opportunity beyond winning accelerator contracts. EPYC can compete for supporting workloads even where the GPUs come from another supplier.
The bullish case is broader than one product launch: more AI workloads create more places for AMD to sell silicon.
A trillion-dollar valuation raises the execution bar. Winning both CPU and GPU deployments is how AMD can grow into it.

[Blue | Semis & AI Infra @BlueTradeIn]

For $MU ’s September 30 earnings, one question deserves more attention:
How much future demand are customers willing to commit to today?
Strong memory prices can produce a spectacular quarter. Longer purchase commitments could make the earnings outlook more durable.
If customers reserve capacity well ahead of delivery, Micron gets better visibility for production planning. If those commitments also protect pricing, the investment case becomes stronger.
That is the bullish development to watch: AI customers treating memory supply as something they need to secure before expanding compute.
Watch contract duration, pricing terms and the cost of adding capacity alongside the headline results.
One great quarter tells you memory is scarce now. Customer commitments help show how long that scarcity might matter.

[Blue | Semis & AI Infra @BlueTradeIn]

A GPU waiting for power and cooling earns its owner nothing.
On September 21, $NVDA launched DSX Ready to qualify power and cooling products for AI factories. $VRT is among the initial cooling suppliers.
The business logic is straightforward: selling faster chips only helps if customers can bring them online.
A clearer set of qualified components can reduce integration risk and help builders turn hardware purchases into working capacity.
For Nvidia, that supports deployment of its platform. For Vertiv, it puts qualified cooling products into the customer’s infrastructure selection process.
Qualification isn’t an order. But being part of the design process matters long before equipment gets purchased.
The AI buildout needs more chips. It also needs fewer reasons those chips sit idle.

[Blue | Semis & AI Infra @BlueTradeIn]

One of AMD’s most important AI upgrades didn’t require a new chip.
In its September 16 MLPerf update, $AMD reported 38% higher GPT-OSS-120B throughput in the Server benchmark using the same eight MI355X GPUs as the previous round, following software optimizations.
That matters to the customer paying for the hardware. More output from the same GPUs can spread the purchase cost across more useful work.
The bullish implication: AMD can make its platform more competitive between chip launches, instead of waiting for the next generation of silicon.
AMD also reported that seven partners closely reproduced its benchmark results. That adds evidence beyond a single reference system.
Benchmarks aren’t purchase orders. But improving performance and repeatability gives customers a stronger reason to evaluate Instinct for their next deployment.
The chip gets sold once. Better software can keep strengthening the case for buying the next one.

[Blue | Semis & AI Infra @BlueTradeIn]

The AI winners are still being debated. The machines needed to build their chips are nearly booked out.
Reuters reported on September 14 that $ASML’s existing EUV machines are almost sold out through 2027, while chipmakers are committing to next-generation High-NA tools. Source
That is the attractive part of this business: ASML doesn’t need to predict which AI model wins. It supplies the manufacturing technology behind competing chip designs.
And the opportunity extends beyond GPUs. Samsung plans to introduce High-NA into DRAM production by 2028, giving ASML another route into AI’s growing memory requirements. Source
The bullish case has two drivers: more manufacturing capacity and more demanding manufacturing technology.
Orders still need to become shipments and profits. But customers committing years ahead is a stronger signal than another optimistic AI forecast.
You can change your favorite AI stock overnight. Building the factories behind it takes years.
$ASML sells into that commitment.

[Blue | Semis & AI Infra @BlueTradeIn]

Joining the S&P 100 won’t sell a single extra SSD.

Sandisk $SNDK enters the index on September 21, replacing Colgate-Palmolive under changes S&P Dow Jones Indices announced September 4.

Two very different buyers matter here: funds buying shares to track an index, and customers buying storage because they need it.

Index-related buying can move the stock around a rebalance. But the date was known weeks ahead. Don’t assume those orders are all still waiting for Monday’s open.

Once portfolios are adjusted, the harder questions remain: are NAND selling prices holding up? Are data-center shipments growing? Is revenue turning into cash?

More bits shipped help, but the gap between selling price and cost per bit determines how much becomes gross profit.

The promotion raises Sandisk’s profile. It doesn’t remove the memory cycle.

The buyer worth watching next is the customer who orders more storage without an index committee telling them to.

[Blue | Semis & AI Infra @BlueTradeIn]

Jensen’s doubling forecast deserves attention. The lazy math doesn’t.
He expects $NVDA to sell twice as many chips next year. That does not automatically mean twice as many AI GPUs or twice the HBM demand.
Nvidia also sells CPUs, networking and other chips. The mix matters.
For $MU investors, the useful calculation is:
AI accelerator shipments × HBM capacity per accelerator × Micron’s share.
If accelerator volumes and memory content both rise, the opportunity compounds. If the extra units are mostly elsewhere, the headline tells you much less about memory.
Supplier qualification, pricing and production yields then determine how much becomes profit.
“Twice as many chips” gets the clicks. Which chips—and whose memory goes into them—determines the earnings.

[Blue | Semis & AI Infra @BlueTradeIn]

A faster model can still be a slow employee.
The GPU finishes generating code. The agent still has to retrieve data, run it, test it and decide what to try next.
$AMD’s September 18 update puts EPYC CPUs directly inside that workflow, with Venice already in production and major cloud deployments expected later this year.
The investment question is changing: how many completed jobs can a server deliver, not just how many tokens can a GPU generate?
CPU-heavy execution can become the bottleneck even when inference gets faster. Buying more GPUs won’t fix that particular queue.
This creates an opening for EPYC, though AMD still has to win deployments against Xeon and Arm alternatives.
The next AI benchmark worth watching: time from “do this” to “done.”
$AMD

[Blue | Semis & AI Infra @BlueTradeIn]

Buying more GPUs is expensive. Buying GPUs that spend their time waiting for data is worse.
$MRVL and $GFS just expanded a multiyear agreement to increase silicon-germanium manufacturing capacity for optical connectivity, covering pluggable transceivers, near-packaged optics and co-packaged optics.
The important detail: this builds on existing production. They’re expanding manufacturing capacity, not just showing a prototype.
As AI workloads spread across more accelerators, moving data between them becomes part of the computing problem. A faster chip cannot fix a congested connection.
That gives Marvell another route into the AI budget beyond designing custom processors: supplying the connectivity that helps expensive clusters do useful work.
For GlobalFoundries, it shows why AI spending can reach specialty manufacturing without every component needing the smallest transistor.
This agreement doesn’t disclose order value or guarantee higher margins. Production ramps and customer adoption still matter.
But the investment implication is worth watching: the networking budget can grow even when customers disagree about which AI accelerator to buy.
$MRVL $GFS

13:21

[Blue | Semis & AI Infra @BlueTradeIn]

The next opportunity for $MU could come from AI companies trying to spend less on HBM.

SK hynix just demonstrated SALT-KV, a system that distributes AI’s cached computations across HBM, DRAM and SSDs based on reuse value and storage cost.
The logic is simple: data that needs immediate access gets expensive memory. Data that can wait gets cheaper capacity.
As agents handle longer conversations and more simultaneous jobs, keeping everything in HBM becomes harder to justify.
That creates an opportunity for Micron across all three layers: HBM, server DRAM and NAND-based SSDs.
The catch matters. Better memory management can reduce capacity needed per task. Total demand grows only if workload growth outweighs those savings. A competitor’s demo also isn’t a Micron order.
But it shows why “AI memory” is a much bigger market than HBM alone.
An SSD doesn’t need to replace HBM to earn a bigger share of the AI budget. It just needs to hold data that was too expensive to keep there.

[Blue | Semis & AI Infra @BlueTradeIn]

Interesting day for NAND.

CXMT wants in.

Solidigm is considering a U.S. fab.

And this is happening while the industry is still dealing with a global memory shortage driven by AI.

The irony:

HBM is so attractive that capex has been pulled toward HBM/DRAM…

…which helped make NAND even tighter.

Now high NAND prices are doing exactly what high prices always do:

They’re attracting supply.

Bullish cycle today.

But never forget what eventually kills a memory cycle:

more fabs.

$MU $SNDK

[Blue | Semis & AI Infra @BlueTradeIn]

AMD can raise prices and still end up collecting someone else’s money.

Reports citing ChannelGate point to roughly 10% chip-price increases in Q4, linked to higher TSMC costs. Treat this as a supply-chain report, not confirmed AMD guidance. Product coverage remains unclear.

The question for $AMD shareholders: how much of the extra revenue survives the supplier invoice?

A simple hypothetical:
Sell a chip for $100, production cost $60 → $40 gross profit.
Raise the price to $110, cost rises to $70 → still $40.
Revenue rises 10%. Gross profit per chip goes nowhere. Gross margin falls from 40% to 36.4%.

But two 10% price hikes don’t automatically cancel out, either. TSMC charges for manufacturing; AMD sells a finished product. Different dollar bases. Wafer costs are only part of AMD’s total cost.

The bullish case is that AMD raises realized selling prices faster than total unit costs, without losing enough volume to erase the gain. Passing through costs is useful. Keeping something extra is better.

For $TSM, the attraction is being harder to replace at the manufacturing step. But its own factories aren’t free: Q2 gross margin was 67.7%, while Q3 guidance was 65–67%. Even the supplier has costs to absorb.

AMD’s baseline: 56% non-GAAP gross margin in Q2, with the same Q3 guidance. That predates the reported Q4 change; it doesn’t prove what the hike will earn.

Watch realized prices, shipment volumes and gross profit together, allowing for product mix. A flat margin on higher sales can still mean more profit.

A bigger invoice makes the headline. Who keeps the extra dollar is the investment story.

[Blue | Semis & AI Infra @BlueTradeIn]

A CPU shortage is only good for Intel if Intel can ship the CPUs profitably.

Seoul Economic Daily reports $INTC CEO Lip-Bu Tan said demand was so strong Intel could serve only about half its customers.

That makes the opportunity clearer. It doesn’t finish the turnaround.

More orders → more shipments → more profit. Each arrow requires execution.

Better factory utilization can spread fixed costs over more chips. Poor yields and expensive capacity additions can eat into the benefit.

Watch shipments, gross margin and cash flow. Strong demand for Intel CPUs also doesn’t automatically validate its external foundry business.

Customers wanting more chips is encouraging. The next earnings reports need to show what Intel earns from supplying them.

[Blue | Semis & AI Infra @BlueTradeIn]

Everyone is debating whether AI demand will slow. The companies renting out the hardware are raising prices.

Reuters reports $NBIS will raise selected on-demand rates from October 1:
$NVDA GPUs: +17–21%
Some CPU-only instances: +25%
Some memory offerings: about +41%

Its second increase in three months.

Better software can make each chip more productive. It doesn’t guarantee that available capacity grows faster than demand.

For $AMD and $INTC investors, the CPU increases deserve attention too. This squeeze extends beyond GPUs.

Cloud rent is not the chipmaker’s selling price. Power, facilities and operating costs matter. The next test is whether customers keep paying the higher rates.

Watch the rental bill, not just the model leaderboard.

[Blue | Semis & AI Infra @BlueTradeIn]

AI’s memory bill is getting so big that old DDR4 deserves a second career.

$ALAB just expanded its Leo memory-controller lineup to support reusing DDR4 and pooling memory across servers.

The problem is simple: one server can run short while another has spare capacity. Buying more DRAM doesn’t fix how poorly the existing memory is shared.

As agents run longer and retain more context, making that capacity accessible becomes more valuable.

This doesn’t turn DDR4 into HBM. Different workloads need different memory tiers.

But it gives Astera another way to earn a place in the AI budget: helping customers use memory they already paid for.

The products are sampling. Volume adoption is the next test.

[Blue | Semis & AI Infra @BlueTradeIn]

Nvidia’s competitors have a frustrating problem: the hardware they’re chasing keeps getting faster after it ships.

$NVDA says software improvements lifted GB300 NVL72 performance by up to 60% on Qwen3-VL between MLPerf benchmark rounds. Same hardware platform. Better execution.

That matters when customers compare a cheaper alternative with the GPUs they already own. The alternative has to justify migration against an improving incumbent, not a frozen spec sheet.

One benchmark doesn’t settle the competition. But it explains why matching Nvidia’s silicon is only part of the job.

The chip ships once. The software team keeps competing.

[Blue | Semis & AI Infra @BlueTradeIn]

The more expensive GPUs get, the more expensive a bad connection becomes.

$CRDO just announced 1.6T optical transceivers with real-time link monitoring. Speed gets the headline. Reliability deserves attention.

In a distributed AI job, a faulty link can leave healthy GPUs waiting. You’re still paying for those chips while they do less useful work.

That gives networking suppliers a powerful sales pitch: protecting the productivity of hardware that costs far more than their own products.

A product launch isn’t a purchase order. Customer adoption and margins still have to follow.

But the investment logic is clear: as AI clusters get bigger, keeping them working together becomes more valuable.