Search podcasts by meaning, not just keywords.
Run a semantic search for a thesis or sector, like "GLP-1 exposure" or "commercial real estate," and read what experts have said about it, speaker-attributed and timestamped.
TOUGHER ECONOMICS
Citing the memory shortage pushing costs higher for GPUs and AI servers, NVIDIA decided to raise prices for all customers by more than 15%.
At the same time, the big three memory producers have little economic incentive to flood the market with enough new capacity to eliminate the shortage.
Why would they?
Margins are near record-highs.
Demand is strong.
Customers are paying (for now.)
Management teams at Micron, Samsung, and SK Hynix refuse to let go of their record-high margins by keeping capacity capex at bay, meaning these shortages are not going away any time soon.
That’s created a split result, naturally expected:
Manufacturers using HBM will raise prices
Competitors will seize the record margins and undercut the big three
Both are already taking place right now.
CXMT and YMCT in China hold a growing share of the HBM market, with plenty of US hardware companies already switching from American and Korean memory to Chinese memory.
It’s like everything else in AI…
Qwen and Kimi are doing a bang-up job at delivering the same outputs that Claude or ChatGPT can, roughly 80% cheaper.
Now memory is coming into the scene, giving cloud computing companies a chance to recover their trillion-dollar capex investment by closing the gap between revenue and costs.
The real alpha from this setup comes by answering the following questions:
How will US/Korean memory be affected?
What does 15% higher prices really mean for hyperscalers and cloud computing?
Is Chinese supply enough to offset all of this
That’s what we’ll cover today.
What happens now that AI just got 15% more expensive?
CHART OF THE DAY
Term premiums on the 10-year Treasury bond have reached a decade high.
What these premiums measure is how much extra compensation investors demand to hold the already “risk-free” instrument in the market.
Because of all the fiscal and monetary policy uncertainty going on right now, these bond investors are unwilling to hold unless they get enough compensation.
Despite of what Bessent does about buybacks, one buyer against hundreds of sellers will do very little to tame the bond markets.
ARE YOU COVERED? —>
10-year Premiums, FRED
In reality, bonds are asking for two simple things:
Inflation stability
Proper economic guidance
Right now, it seems like both the Fed and the Treasury are choosing to kick the can down the road as long as the AI financing remains untouched.
Go figure, if we end the AI mania, inflation and fiscal uncertainty likely gets fixed.
THE TRUTH OF PRICE HIKES
AI Server Bill of Materials, Offside Capital
I’ve kept up with the sentiment around the NVIDIA announcement over the weekend, and after a few hundred surveys, here’s what the prevalent read is:
NVIDIA has so much demand, that they have no choice but to raise prices
This hike will send DRAM memory names to new all-time highs
The reality couldn’t be further from this.
I’ve seen enough industry cycles in my career to tell you one thing…
Companies don’t raise prices when they’re at the peak of their earnings wave, they raise them when their margins are thinning beyond a very specific threshold.
Because HBM now makes up ~12% more of the overall cost stack for NVIDIA, and expected to rise by another ~10% by next year, the company’s margins are going to be significantly squeezed.
If that happens, then the hundreds of entities they have bought equity stakes in will lose value, and covenants on purchase agreements will also go up in smoke.
Simply put, they cannot afford to lose the EPS momentum, even if it means cutting some customers at the knees by hiking prices.
AI Trade Profit Pool, Offside Capital
In my memory capital cycle peak analysis, I showed you how these peak margin episodes tend to drive in a new wave of competitors to undercut those who refuse to listen to the market.
NVIDIA’s price hikes is simply a choice to postpone the inevitable:
Raise prices to keep buying Micron chips, so Micron can afford to pay Oracle, who can then afford to pay OpenAI
Why not just choose a different HBM supplier?
After all, Chinese competition has proven to make high-quality chips at a fraction of the cost.
Because if they choose a supplier outside of the big three, then the entire circular financing scheme goes down the drain.
Why do I say this with such confidence?
Remove the circular financing commitments, and this is what you start to see in the free market:
Chinese Memory Orders from US Companies, Offside Capital
HP, Dell, and the big one… Apple.
They’ve all moved to adopt Chinese memory chips from CXMT and YMCT alike for their hardware products.
Why?
Because it’s the logical choice, they get just as good a product for a much lower cost and greedy contract pricing.
Remember the billion-dollar contracts Micron was touting last month?
I suspect a lot of those promises will go unfulfilled as more customers choose to protect their bottom line and shareholder returns.
Caveat being: So long as they are not exposed to NVIDIA’s circular financing, which essentially requires American memory makers to continue pushing higher.
ROI CHALLENGE
Monetized AI Economics, Offside Capital
Breaking down the cost structure and subsequent ROI requirements will show a very dangerous shift to everyone involved.
Especially the hyperscalers…
With a 15% price hike, the required ROIC floor goes from ~11.5% to ~12.5%.
10-year bond yields trading at 4.7% will boost this ROIC to a floor of ~17.5% for every dollar of capex these big spenders deploy.
In other words,
An already difficult path to monetization was just made a lot harder to achieve thanks to NVIDIA.
I will repeat myself in hopes that it all sticks.
Hyperscalers can easily shift to Chinese memory to protect their margins and debt burdens
Doing so would undercut American semis/memory players
That would cut the payments being made to OpenAI and Anthropic
As we already covered, Microsoft derives ~80% of its AI revenue from OpenAI.
More than 60% of Amazon and Google EPS come from unrealized equity gains in their Anthropic stakes.
Simply put,
They won’t change their supplier because that would end up adjusting their earnings lower by 60% and crash the entire AI party.
Now you understand why Anthropic is rushing to have its IPO as soon as humanly possible.
Spend vs Returns - Depreciation Boosts, Offside Capital
Right now, hypercalers receive $1 back for every $13 they spend on AI capex.
After the 15% hikes and the new ROIC hurdle rate, this will look more like $25 of spend for $1 of returns.
Which is exactly where the depreciation schedule extension comes to save the day.
As Chinese competition undercuts the big three memory players, new AI servers and GPUs will lose value faster than anyone thought before.
Therefore,
Depreciation charges for these hyperscalers and cloud compute players could double on average, combined with equity stake drawdowns hurting bottom-line EPS further.
I wonder if this is why capital has rotated from AI into Bitcoin this week…
I don’t yet think the market is aware of this mess, because it gets rather complex in a very quick way.
Still,
The laws of capital never change, no matter how neatly you package them.
Which is why I chose memory as my short candidate, and executed this long/short put option spread on the most exposed name of all.
That trade is already up 30% in a week, and I expect a huge payoff by the fourth quarter of 2026.
WHAT’S THE TRADE?
The best gauge I’ve found to follow the AI trade is the SMH ETF. If you’re not yet part of Offside Premium, I would recommend you abstain from trading individual AI stocks by yourself.
You can still follow the ETF though, as I announced the failed breakout at $591 a couple of weeks ago.
SMH ETF, Thinkorswim
It’s now down ~8% since I gave my initial signal, and if we fail to see a massive rebound out of ~$535, I suspect a new bear market may be upon the AI trade.
The reason is simple,
20% off the highs is the proverbial “bear market” threshold, and mechanically this is where leverage gets called.
Meaning, participants need to either liquidate their holdings or post collateral.
Hedge funds already closed off most of their AI exposure at the failed breakout according to Goldman Sachs prime brokerage data (proprietary of Offside Capital.)
I suspect they did this ahead of that 20% level margin call, as most other participants are short on cash and highly levered.
This idea could 100% be proven wrong if no liquidation comes, but that means Offside Premium members are sitting comfortably with enough downside cushion to wait for that to happen.
Show me a better ROI.
A Final Note
COMING UP NEXT
There have been enough macro movements this quarter to warrant a proper macro update for you
Recent data out of GDP, labor, and business are pointing in a new direction for the AI trade
Gold, USD, and yields seem to have shifted toward fiscal unease, kickstarting the “debasement” trade once again
The Maverick of Wall Street just updated his take on the market’s brand new financing scheme, covering the hidden obligations carried off-balance sheet for most of the hyperscalers.
Rising bond yields are quite honestly trying to put a stop to what the Fed and Treasury seem to be oblivious to:
Until next time,
OFFSIDE RESEARCH
Against the Tape, Ahead of the Curve.