M
monica
Guest
AI's money is flowing back up the supply chain. Samsung posts record profit, up 782%. OpenAI's revenue shrinks by $20 billion.
Let me start with a number that felt almost unreal.
Samsung Electronics just posted quarterly operating profit of 107.4 trillion won, roughly $80.2 billion. It is the first time in the company's nearly sixty-year history that quarterly profit broke 100 trillion won. Year over year, that is up 782%.
The next day, Samsung's stock fell 0.7%.
The drop was tiny, almost negligible. What is interesting is the contrast: a company delivers the best result in its history, and the market's response is a barely perceptible decline.
Analyst Josh Gilbert put it more bluntly: record profit, and still short of expectations. That, he said, tells you how demanding the AI trade has become. [1]
The same week, OpenAI's annualized revenue was revised down from a previously reported $70 billion to "approaching $50 billion." A $20 billion gap. [2]
It reminded me of a very old metaphor. AI is a gold rush, and the people who actually get rich are never the ones digging for gold.
It's the ones selling shovels.
Start with Samsung.
The core of this profit is HBM, high-bandwidth memory, the most sought-after component in AI servers. Every GPU has to be surrounded by it. And that market is effectively a duopoly between Samsung and SK Hynix. Every GPU Nvidia sells, these two take a cut.
This is not the first time Samsung has surprised the market. Last quarter it posted record revenue of 171.5 trillion won and record operating profit of 89.5 trillion won. [1] This quarter, it stepped up again. Two consecutive record quarters is not luck; it is demand pushing from underneath.
The telling change is that memory buyers are now signing multi-year supply agreements.
In a business famous for brutal cycles, where investors always dread the next downturn, buyers locking in supply for years means only one thing: they are afraid they won't get supply at all.
So in this cycle, memory has changed category. It moved from a cyclical commodity to a strategic resource. Commodity prices swing with the cycle. Strategic-resource prices climb with your competitor's anxiety.
Then the chip layer.
Nvidia announced $1 billion over five years for U.S. superintelligence and quantum research this week. [3] A chip company donating money to science is itself a signal that its cash pile has outgrown its own needs.
The company that made me realize how overheated this has become is one you may not know: Etched. It builds AI inference chips aimed squarely at Nvidia. Two months ago it raised $700 million at a $21 billion valuation. Two months later, investors are reportedly offering terms at $40 to $50 billion. [4]
The details are more telling. About 15% of this 400-person company came from Nvidia. Its last round was led by quant giant Jane Street, which didn't just write a check, it bought Etched's systems and became a customer. Etched says it has $1 billion in orders.
Etched's playbook deserves a note. It doesn't go head-to-head with Nvidia on general-purpose GPUs. It focuses on inference, the step where a model spits out a response every time you prompt it. Its pitch is one sentence: for the same dollar, we give you more tokens. For a quant firm, that is lethal, because a few tenths of a millisecond can be real money.
Behind this is a bigger shift. AI chip competition is moving from training to inference. Training is a one-time heavy spend; inference is the burn that happens every second of every day. Whoever drives down the cost of inference holds the pricing power of the next phase.
The third link is quieter, but may be the most certain winner: power.
AI data centers consume so much electricity that the U.S. government is drafting a law specifically about them. A bipartisan Senate bill would require data centers to report their power use and let utilities charge them more for transmission. [5]
Notice that signal. Charging an industry extra for electricity is almost unheard of. It means policymakers have already reclassified data centers from "high-tech" to "power hogs."
States have moved too. New York paused large data centers; New Jersey residents are protesting. [6] To locals, a data center is a neighbor that grabs power, dodges taxes, and makes noise.
None of it stops them from being built.
Everyone talks about models, chips, parameters. Few register that a single large data center can draw as much power as a mid-sized city. Electricity scarcity is becoming the physical ceiling of AI expansion, and anything that touches a ceiling gets repriced.
That's why I call power the most certain winner: no matter whose model or whose chip you end up using, you still have to pay the wire.
Now, downstream, the model makers.
A little over a week ago, reports said OpenAI's annualized revenue was nearing $70 billion, enough to rival Anthropic and enough to justify one massive round after another. Then the Financial Times reported that OpenAI told investors the real number is "approaching $50 billion."
The interesting part is where that $70 billion came from. Per the FT, it was an artifact of OpenAI's own investors trying to build a direct comparison with Anthropic. Anthropic counts sales made by its cloud partners in its run rate; OpenAI doesn't.
So the "shrinkage" is, at bottom, a collapse of an accounting convention.
But the market doesn't care about conventions. It sees the other facts: an IPO pushed back to early 2027; a March funding round that alone raised $122 billion; leaked 2025 financials showing about $13 billion in revenue against far higher spending.
Making models doesn't make money yet. What it makes is nowhere near the hole it digs.
OpenAI is not alone. Its rival Anthropic is the same question written a different way.
Anthropic's annualized revenue climbed to $65 billion in July, growth is fast, and the market is rewarding it, reportedly gearing up for an IPO this fall with valuations floated above $2 trillion. Its CEO, Dario Amodei, took home $18 million in 2025, higher than Google's Sundar Pichai or Amazon's Andy Jassy. [7]
But revenue, however fast, can't outrun the spend. The model-layer arms race is a burn game: training burns, inference burns, talent burns, compute burns the most.
So you get a telling structure: two downstream giants, one shrinking by $20 billion, the other sprinting toward a $2 trillion valuation. Different fortunes, same fact.
Put upstream and downstream side by side, and an asymmetry appears that is rarely discussed.
Downstream sells promises. OpenAI's annualized revenue and Anthropic's run rate are, at bottom, projections of future income, a check that hasn't fully cleared. They use that check to raise money, and the money to buy what the upstream makes.
Upstream sells goods. Samsung's memory, Nvidia's GPUs, the grid's power: delivered today, paid for today. An order is an order, cash is cash.
When a cycle's narrative starts to strain, the market doubts the uncashed check first, not the cash already in hand.
That is why money flows downstream to upstream. It isn't a question of who is smarter. It's the underlying structure of the two businesses.
There is a deeper shift, too, about moats.
For two decades, the dominant narrative was "software eats the world." Software moats come from network effects, data, user lock-in. Hardware and infrastructure were seen as low-margin, heavy-asset drudgery.
This AI wave has inverted that ranking.
A model's lead is catchable. Talent moves, open source spreads, and training gaps converge over time. Lead by a year today, and you may be caught tomorrow.
An upstream lead is physical. TSMC's advanced nodes, HBM capacity ramps, grid expansion: these can't be replicated by throwing money at them. Their build cycles run in decades, and their moats are carved into the constraints of the physical world.
So the genuinely interesting thing about this wave is that the fruits of the technological revolution are flowing, for the first time, mainly into manufacturing, energy, and infrastructure, rather than software and applications. It is a quiet reversal of the last two decades.
Every boom of this kind has followed roughly the same script.
Downstream tells the story, upstream takes the orders; downstream raises, upstream expands; then downstream's story breaks, and upstream's capacity turns into oversupply. The 2000 internet bubble and the 2017 mining boom played out this way.
But this time there is a variable that may change the ending.
In the past, upstream expanded fast. Chips, memory, and power have all lived through oversupply. This time is different, because AI doesn't need ordinary chips and memory. It needs HBM, advanced nodes, liquid cooling, and grid capacity. These expand slowly, and their barriers are extreme.
So the more likely script this time isn't "upstream crashes along with downstream." It's "downstream consolidates, upstream oligopolizes." A few downstream players die, a few upstream players remain, and the ones that remain get more expensive in the next cycle.
If these judgments hold, then the thing worth watching is not Samsung's earnings. It's two other questions.
First, whether the model companies can make revenue growth outpace their burn before the funding window closes. OpenAI's shrunken $20 billion, its delayed IPO, Anthropic's $2 trillion tightrope: these are different sides of the same question.
Second, how much patience the market still has for the AGI narrative. Because every dollar of upstream profit ultimately comes from downstream's ability to tell a story. When the drum stops, the potato stops moving.
When shovel sellers post 782% growth and the market still finds "a record" not enough, expectations have been pushed to a dangerous height. Samsung is probably just the first. Who's next to fail to deliver "good enough"?
AISemiconductorsSupply ChainOpenAISamsung
Let me start with a number that felt almost unreal.
Samsung Electronics just posted quarterly operating profit of 107.4 trillion won, roughly $80.2 billion. It is the first time in the company's nearly sixty-year history that quarterly profit broke 100 trillion won. Year over year, that is up 782%.
The next day, Samsung's stock fell 0.7%.
The drop was tiny, almost negligible. What is interesting is the contrast: a company delivers the best result in its history, and the market's response is a barely perceptible decline.
Analyst Josh Gilbert put it more bluntly: record profit, and still short of expectations. That, he said, tells you how demanding the AI trade has become. [1]
The same week, OpenAI's annualized revenue was revised down from a previously reported $70 billion to "approaching $50 billion." A $20 billion gap. [2]
It reminded me of a very old metaphor. AI is a gold rush, and the people who actually get rich are never the ones digging for gold.
It's the ones selling shovels.
Memory became a strategic resource
Start with Samsung.
The core of this profit is HBM, high-bandwidth memory, the most sought-after component in AI servers. Every GPU has to be surrounded by it. And that market is effectively a duopoly between Samsung and SK Hynix. Every GPU Nvidia sells, these two take a cut.
This is not the first time Samsung has surprised the market. Last quarter it posted record revenue of 171.5 trillion won and record operating profit of 89.5 trillion won. [1] This quarter, it stepped up again. Two consecutive record quarters is not luck; it is demand pushing from underneath.
The telling change is that memory buyers are now signing multi-year supply agreements.
In a business famous for brutal cycles, where investors always dread the next downturn, buyers locking in supply for years means only one thing: they are afraid they won't get supply at all.
So in this cycle, memory has changed category. It moved from a cyclical commodity to a strategic resource. Commodity prices swing with the cycle. Strategic-resource prices climb with your competitor's anxiety.
Chips are shifting from training to inference
Then the chip layer.
Nvidia announced $1 billion over five years for U.S. superintelligence and quantum research this week. [3] A chip company donating money to science is itself a signal that its cash pile has outgrown its own needs.
The company that made me realize how overheated this has become is one you may not know: Etched. It builds AI inference chips aimed squarely at Nvidia. Two months ago it raised $700 million at a $21 billion valuation. Two months later, investors are reportedly offering terms at $40 to $50 billion. [4]
The details are more telling. About 15% of this 400-person company came from Nvidia. Its last round was led by quant giant Jane Street, which didn't just write a check, it bought Etched's systems and became a customer. Etched says it has $1 billion in orders.
Etched's playbook deserves a note. It doesn't go head-to-head with Nvidia on general-purpose GPUs. It focuses on inference, the step where a model spits out a response every time you prompt it. Its pitch is one sentence: for the same dollar, we give you more tokens. For a quant firm, that is lethal, because a few tenths of a millisecond can be real money.
Behind this is a bigger shift. AI chip competition is moving from training to inference. Training is a one-time heavy spend; inference is the burn that happens every second of every day. Whoever drives down the cost of inference holds the pricing power of the next phase.
Power, the underestimated link
The third link is quieter, but may be the most certain winner: power.
AI data centers consume so much electricity that the U.S. government is drafting a law specifically about them. A bipartisan Senate bill would require data centers to report their power use and let utilities charge them more for transmission. [5]
Notice that signal. Charging an industry extra for electricity is almost unheard of. It means policymakers have already reclassified data centers from "high-tech" to "power hogs."
States have moved too. New York paused large data centers; New Jersey residents are protesting. [6] To locals, a data center is a neighbor that grabs power, dodges taxes, and makes noise.
None of it stops them from being built.
Everyone talks about models, chips, parameters. Few register that a single large data center can draw as much power as a mid-sized city. Electricity scarcity is becoming the physical ceiling of AI expansion, and anything that touches a ceiling gets repriced.
That's why I call power the most certain winner: no matter whose model or whose chip you end up using, you still have to pay the wire.
The model makers are still filling a deep hole
Now, downstream, the model makers.
A little over a week ago, reports said OpenAI's annualized revenue was nearing $70 billion, enough to rival Anthropic and enough to justify one massive round after another. Then the Financial Times reported that OpenAI told investors the real number is "approaching $50 billion."
The interesting part is where that $70 billion came from. Per the FT, it was an artifact of OpenAI's own investors trying to build a direct comparison with Anthropic. Anthropic counts sales made by its cloud partners in its run rate; OpenAI doesn't.
So the "shrinkage" is, at bottom, a collapse of an accounting convention.
But the market doesn't care about conventions. It sees the other facts: an IPO pushed back to early 2027; a March funding round that alone raised $122 billion; leaked 2025 financials showing about $13 billion in revenue against far higher spending.
Making models doesn't make money yet. What it makes is nowhere near the hole it digs.
OpenAI is not alone. Its rival Anthropic is the same question written a different way.
Anthropic's annualized revenue climbed to $65 billion in July, growth is fast, and the market is rewarding it, reportedly gearing up for an IPO this fall with valuations floated above $2 trillion. Its CEO, Dario Amodei, took home $18 million in 2025, higher than Google's Sundar Pichai or Amazon's Andy Jassy. [7]
But revenue, however fast, can't outrun the spend. The model-layer arms race is a burn game: training burns, inference burns, talent burns, compute burns the most.
So you get a telling structure: two downstream giants, one shrinking by $20 billion, the other sprinting toward a $2 trillion valuation. Different fortunes, same fact.
Downstream sells promises, upstream sells goods
Put upstream and downstream side by side, and an asymmetry appears that is rarely discussed.
Downstream sells promises. OpenAI's annualized revenue and Anthropic's run rate are, at bottom, projections of future income, a check that hasn't fully cleared. They use that check to raise money, and the money to buy what the upstream makes.
Upstream sells goods. Samsung's memory, Nvidia's GPUs, the grid's power: delivered today, paid for today. An order is an order, cash is cash.
When a cycle's narrative starts to strain, the market doubts the uncashed check first, not the cash already in hand.
That is why money flows downstream to upstream. It isn't a question of who is smarter. It's the underlying structure of the two businesses.
The moat got quietly re-ranked
There is a deeper shift, too, about moats.
For two decades, the dominant narrative was "software eats the world." Software moats come from network effects, data, user lock-in. Hardware and infrastructure were seen as low-margin, heavy-asset drudgery.
This AI wave has inverted that ranking.
A model's lead is catchable. Talent moves, open source spreads, and training gaps converge over time. Lead by a year today, and you may be caught tomorrow.
An upstream lead is physical. TSMC's advanced nodes, HBM capacity ramps, grid expansion: these can't be replicated by throwing money at them. Their build cycles run in decades, and their moats are carved into the constraints of the physical world.
So the genuinely interesting thing about this wave is that the fruits of the technological revolution are flowing, for the first time, mainly into manufacturing, energy, and infrastructure, rather than software and applications. It is a quiet reversal of the last two decades.
History's script, and this time's variable
Every boom of this kind has followed roughly the same script.
Downstream tells the story, upstream takes the orders; downstream raises, upstream expands; then downstream's story breaks, and upstream's capacity turns into oversupply. The 2000 internet bubble and the 2017 mining boom played out this way.
But this time there is a variable that may change the ending.
In the past, upstream expanded fast. Chips, memory, and power have all lived through oversupply. This time is different, because AI doesn't need ordinary chips and memory. It needs HBM, advanced nodes, liquid cooling, and grid capacity. These expand slowly, and their barriers are extreme.
So the more likely script this time isn't "upstream crashes along with downstream." It's "downstream consolidates, upstream oligopolizes." A few downstream players die, a few upstream players remain, and the ones that remain get more expensive in the next cycle.
When the drum stops, the potato stops moving
If these judgments hold, then the thing worth watching is not Samsung's earnings. It's two other questions.
First, whether the model companies can make revenue growth outpace their burn before the funding window closes. OpenAI's shrunken $20 billion, its delayed IPO, Anthropic's $2 trillion tightrope: these are different sides of the same question.
Second, how much patience the market still has for the AGI narrative. Because every dollar of upstream profit ultimately comes from downstream's ability to tell a story. When the drum stops, the potato stops moving.
When shovel sellers post 782% growth and the market still finds "a record" not enough, expectations have been pushed to a dangerous height. Samsung is probably just the first. Who's next to fail to deliver "good enough"?
Sources
- 1. Samsung forecasts record third-quarter profit of $80 billion on AI boom — CNBC
- 2. OpenAI's revenue is reportedly $20 billion less than previously projected — TechCrunch
- 3. Nvidia 'commits' $1 billion to US science research — The Verge
- 4. Etched fields funding offers at $40B+ valuation, sources say — TechCrunch
- 5. A bipartisan bill aims to hold data centers accountable for their energy use — The Verge
- 6. This week in the big AI data center buildout — The Verge
- 7. Dario Amodei earned $18 million as Anthropic's CEO in 2025 — The Verge
AISemiconductorsSupply ChainOpenAISamsung