The AI Stack Just Split in Half
The Headline Everyone Ran
The coverage of Kimi K3 centered on the spectacle: a Chinese AI model designed a chip without human help. Cadence and Synopsys shares cratered. Analysts scrambled to assess whether the EDA duopoly faces disruption. BNP Paribas told clients to buy the dip. Morgan Stanley framed it as cumulative progress across China’s model industry, not an overnight extinction event.
Both takes are defensible. Neither asks the question that matters for capital allocation: which layer of the AI infrastructure stack has a moat, and which one just discovered it might not?
Two Weeks, Two Software Casualties
The tape has been marking this territory for two weeks running. On July 15, IBM fell 20% because its enterprise clients redirected budgets from software licenses to AI servers and memory. That was a demand signal: the money moved from the code layer to the hardware layer. Friday, Kimi K3 demonstrated the same migration in engineering terms—an AI replaced licensed chip-design software with an open-source stack.
The bond market did not flinch. The 10-year yield is falling, not rising. In this tape, that quiet confirms the read: the AI infrastructure build is not a macro overheating story. It is a reallocation story. Capital is flowing down the stack, from the services and tools layer toward the physical substrate—the fabs, the memory, the power.
Brent crude at $88 on the back of a 14% weekly gain adds a second confirmation. The economy is repricing physical scarcity, not digital leverage. Our view: when two consecutive weeks produce two different software casualties through two different mechanisms, the signal is structural, not anecdotal.
Where the Margin Goes When AI Designs Its Own Tools
The investment question is not whether Kimi K3 can replace Synopsys at the frontier 3nm node today. It cannot. The demo ran at 45nm, several generations behind. The question is the direction of travel and the pricing-power shift it implies over the next six to eighteen months.
EDA companies have traded at premium multiples for decades because their tools sit between every chip designer and every manufactured wafer. Cadence trades at roughly 41 times forward earnings. Synopsys at 24 times. Those multiples price a durable moat. If frontier AI models move down the capability curve in chip design the way they have in coding, legal research, and image generation, the moat does not vanish overnight. But it narrows. And the market has started repricing the toll.
Meanwhile, the companies that own the physical infrastructure—the fabs, the packaging, the power plants—face no open-source competitor. You cannot download a chip foundry. In this tape, hardware scarcity is the moat that holds.
Worth watching: semiconductor revenue grew 79% year-on-year in Q1 2026, per BNP Paribas. They expect Q2 growth to accelerate to 132%. The money is not leaving the AI trade. It is moving to a different floor of the building.
Wednesday Settles the Bet
Alphabet reports Wednesday after the close. Its Google Cloud business is the largest single buyer of custom silicon in the AI stack, and the company designs its own TPU chips in-house—exactly the kind of buyer that benefits if design-tool costs fall. Tesla reports the same evening, per CNBC. Its capital-expenditure tripling into robotics and AI shows where the money is going: physical machines, not software layers.
Intel reports Thursday. Its 160% gain in 2026 is the hardware-cycle bet in its purest form.
The tripwire: if Alphabet signals accelerating custom-chip investment while guiding cloud capex higher, the market will confirm that AI is good for the companies that buy design tools and bad for the companies that sell them. If Intel delivers and guides above consensus, the hardware-over-software rotation gets another floor under it. If either disappoints, Friday’s EDA selloff becomes an overreaction and the moat question resets. Wednesday evening is the verdict.
|