Mind the Spread

Tracking second-order stock ideas across AI, life sciences, robotics, energy and market structure. Notes only, not financial advice.

[Mind the Spread @Mind_The_Spread]

@sama Can you reset codex limits for plus users please and thanks.

[Mind the Spread @Mind_The_Spread]

$FDS is an awkward sort of AI trade.

Its long-term moat may be narrowing. Its near-term product may be becoming more useful.

Both can be true.

If clean financial data becomes easier to sell through agent connectors, seat growth could improve well before the terminal debate is settled.

I hold a position. Watching the numbers, not the label.

[Mind the Spread @Mind_The_Spread]

I’ve started a small position in $ALGM.

The market still sees an automotive sensor business waiting for the cycle to turn.

The more interesting possibility is robotics, where machines need precise sensing to know what has moved and whether contact has been made.

The motor trade pays the bills. Robotics provides the option.

Probably early. That is rather the point.

$ALGM

[Citrini @citrini]

The selloff has been centered, thus far, on the most obvious beneficiaries of AI capex spending. Over the weekend, we decided to revisit our attempts to discover companies with real upside that are still less established as consensus AI trades.

https://t.co/UxucfP9NBa

[Mind the Spread @Mind_The_Spread]

@jukan05 @deanwball This is the only way decoupling works.

By making procurement, compliance and reputational risk awkward enough, companies will just quietly step away.

The winners will be the US open source ecosystem and the security, governance and infrastructure needed to deploy it properly.

[Mind the Spread @Mind_The_Spread]

I’ve added $META.

The advert business still funds the AI build, but the more interesting question is whether that infrastructure becomes a revenue stream in its own right.

If Meta can monetise spare compute as well as improve adverts, today’s capex looks rather less indulgent.

Execution risk remains. The optionality is improving.

[Mind the Spread @Mind_The_Spread]

$ISRG beat on revenue and earnings.

The shares fell 14%.

Rather neatly sums up the current tape.

The company delivered. The valuation had already demanded more.

At a heroic multiple, a beat is merely the entry requirement.

The question is whether the de-rating has now moved faster than the earnings downgrade.

[Mind the Spread @Mind_The_Spread]

$ARM has been treated rather roughly this week.

The easy conclusion is that the AI trade has cracked.

The more useful conclusion is that positioning had become far too one-way.

Arm still sits across:

• mobile
• edge AI
• robotics
• data centres
• custom silicon

None of that makes the shares cheap.

It does, however, mean the investment case has not deteriorated at anything like the pace of the multiple.

That distinction matters.

In a de-grossing event, the first question is not:

“What has fallen the most?”

It is:

“Where are earnings revisions still moving higher once the forced selling has finished?”

A broken chart is not always a broken company.

Sometimes the book was simply too crowded.

[Mind the Spread @Mind_The_Spread]

$TSM just delivered the kind of quarter semiconductor bulls dream about.

Revenue up 36%.
Profit up 77%.
Record margins.
Guidance above expectations.

The stock fell anyway.

That does not mean AI demand is broken.

It means expectations have become exceptional.

When a trade gets crowded, strong results stop being a catalyst.

They become the minimum requirement.

That is the second read on this week’s semiconductor sell off:

The fundamentals can remain intact while the price gets punished for positioning, valuation and everyone already owning the same idea.

The best company can still be the wrong trade at the wrong price.

But violent de-risking can also create the next setup.

The question now is not who beat estimates.

It is who still has revisions left once the crowd has gone.

[Mind the Spread @Mind_The_Spread]

$SNDK has probably been one of the more interesting red candle names this week.

The obvious take:

Memory got too hot.
Crowding unwinds.
Sell the whole group

But the second read is more interesting.

If AI inference keeps running into the memory wall, the answer may not just be “buy more HBM forever”.

It may be a new memory hierarchy.

HBM for speed.
Flash for capacity.
Controllers for intelligence.
Architecture doing more of the work.

That is where HBF gets interesting.

$SNDK is not just a cyclical NAND name if HBF becomes a serious tier between HBM and SSDs.

The stock can be overextended.
The selloff can be deserved.
And the long-term setup can still be real.

All three can be true.

The question is whether this is a broken memory trade…

Or just the market violently repricing one of the few credible ways around the DRAM tax.

[Mind the Spread @Mind_The_Spread]

The market still treats AGI like a faster software cycle.

The people closest to the frontier are increasingly talking like it is a new regulatory and infrastructure regime.

Whether the timeline is right or wrong, the second order read is clear:

evals, identity, security, observability, compute access and governance all become much more important layers.

[Mind the Spread @Mind_The_Spread]

Everyone is watching AI.

But oil might be the thing that decides what AI stocks are worth next.

If Hormuz stays hot, this is not just an energy headline.

It hits:

• Inflation
• Freight
• Chemicals
• Power
• Rates
• Margins
• Consumer Spend

That matters because the AI trade is priced like the world stays calm.

Cheap capital. Stable inputs. Endless capex. No macro interruption.

Maybe that is right.

But if oil starts acting like a real bottleneck again, the market may have to remember something uncomfortable:

The digital economy still sits on top of the physical one.

Data centres need power. Power needs fuel. Supply chains need shipping. Consumers need spare cash.

AI can be structurally right and still get repriced by oil.

That is the bit I think people are underweighting.

What breaks first if energy keeps squeezing?

Margins, multiples, or the soft-landing narrative?

[Mind the Spread @Mind_The_Spread]

$OUST is one of the cleaner ways to play robotics without betting on which humanoid wins.

The obvious take:

Robots need a brain. Buy compute. $$$

But physical AI also needs eyes, chuck.

A robot that can’t perceive depth, motion, colour and objects in real time is just expensive metal with a GPU strapped to it.

That is why colour LiDAR is interesting.

It starts to collapse the old camera vs LiDAR debate into one sensor stack:

• Depth
• Colour
• Timing
• Localisation
• Object detection
• Cleaner training data

For robots, latency is not a feature.

It is survival.

$OUST is still speculative, still volatile, and definitely not the whole robotics trade.

But if robotics starts moving from demo videos and kicking kids in the head to actual deployment, perception is one of the first layers I’d expect the market to care about.

[Mind the Spread @Mind_The_Spread]

@teortaxesTex hackernews just points out to media slop or old twitter news and the comment section has been deranged for at least a couple of years.

It was pretty useful during the SAAS years but a post hasn't licked my favourite bar in a while.

[Mind the Spread @Mind_The_Spread]

@zephyr_z9 Chat Pro = give me the best answer right just now.

Work Max/Ultra = bugger off, use tools/files/apps, coordinate steps and then come back when you're finished.

Same brain family. Different gearbox.

[Mind the Spread @Mind_The_Spread]

The memory shortage is not just a semi problem.

It is a tax on the entire AI stack.

$NVDA can juggle compute, networking and memory because the economics justify it.

Hyperscalers get efficiency gains and spend them immediately on more compute.

$AAPL does not have that luxury and is staring down the barrel of a gun!

Same scarce memory pool.
Very different pricing power.
If AI servers keep outbidding consumer devices, the iPhone BOM gets ugly fast.

Jevons for hyperscalers.
Margin squeeze for some, miniature American flags for others!

[Mind the Spread @Mind_The_Spread]

This is the key distinction for me.

AI demand can be structurally real, while memory still behaves cyclically.

Those two things are not contradictory.

TSMC can stay relatively disciplined because of the customer relationship.

Memory is where the stress gets expressed most violently because pricing clears much closer to spot supply/demand.

So the read through is not “the cycle is dead”.

AI has made the slope steeper and the peak potentially higher.

But if DRAM/HBM margins get pushed into absurd territory, the eventual mean reversion will probably be just as violent.

Great trade, dangerous religion.

[Mind the Spread @Mind_The_Spread]

$SDGR is one of those names where the first take and second take are very different.

First take:

AI drug discovery has been hyped for years.
Biotech funding has been awful.
Most drugs still fail.
Move on.
Fair.
But the second read is more interesting.
AI doesn’t need to “cure cancer” tomorrow for this to matter.

It just needs to improve the hit rate.

More credible programmes entering validation.
Bad hypotheses killed earlier.
Capital recycled faster.
Better preclinical evidence before pharma writes the cheque.

That changes the maths.

$SDGR sits in the speculative TechBio bucket, so this is not the same as owning the picks-and-shovels around trials, manufacturing or bioprocessing.

But if the market starts believing AI can improve R&D velocity, not just generate pretty molecules, names like this become worth watching again.
The setup I’m watching:
• biotech funding thaw
• pharma patent cliff
• AI-originated trials
• faster preclinical filtering
• M&A appetite returning
• sentiment still washed out
Not financial advice.

Just one where the market may have thrown the useful bit out with the hype.

[Mind the Spread @Mind_The_Spread]

$FDS is one of the more interesting “AI loser” names.

The lazy take is simple:

AI eats terminals. Analysts need fewer seats. Data vendors lose.

Maybe.

But the second read is that agents still need clean, permissioned, reliable financial data.

If your AI workflow can write the model, screen the market and draft the note, the bottleneck becomes the data pipe.

Bad inputs, bad outputs.

That is where FactSet gets interesting.

Not because it is suddenly an AI hypergrowth name.

Because it may become useful plumbing for AI-native finance workflows while the market still treats it like old-world terminal spend.

The setup I’m watching:

• MCP/connectors
• New workflow seats
• Cleaner financial data
• AI tools needing trusted sources

Just one where consensus might have gone a bit too neat.