Goldman Sachs summarizes the CES analyst conference: The incremental demand for AI infrastructure comes from "embodied intelligence" and "agents"

Don't be disturbed by the noise in the market about "the peak of the AI cycle." Goldman Sachs held intensive meetings with semiconductor giants such as NVIDIA, AMD, ADI, Marvell, and Micron during CES 2026, revealing a deeper truth: this is not a simple cyclical cycle, but a deep evolution of the AI infrastructure frenzy.

According to the Wind Trading Desk, on January 7th, Goldman Sachs pointed out in its latest research report: the demand for AI infrastructure remains strong, but the driving forces are undergoing subtle structural changes—the mid-term increments will mainly come from "physical AI" (i.e., embodied intelligence) and "agent AI."

For investors, this means that merely focusing on computing power stacking is no longer sufficient; attention must be paid to the hardware evolution that can support longer contexts and more complex reasoning capabilities. NVIDIA is not just selling graphics cards; it is redefining memory hierarchy through new storage platforms, directly benefiting NAND demand. Meanwhile, although inventory in the analog chip sector is extremely low, OEM manufacturers have not yet begun large-scale inventory replenishment, and the recovery shows characteristics of an "L-shaped" bottom.

NVIDIA: The Rubin platform is not just computing power, but a redefinition of "memory"

Jensen Huang's empire has not only confirmed the durability of AI infrastructure demand but also demonstrated its absolute control over the next generation of computing forms. The highly anticipated Rubin platform is set to see strong production ramp-up in the second half of 2026 (2H26). Here is a shocking production detail: the assembly time for the Rubin tray is only about 5 minutes, whereas the Blackwell tray takes about 2 hours. This leap in production efficiency means that supply bottlenecks will be completely broken, and management has clearly stated that there are currently no major supply constraints for GB200/300.

But the real "trump card" is about storage. To accelerate the landing of "agent AI," NVIDIA has launched a new platform that utilizes SSDs to expand working context memory. Previously, each GPU could only access about 1TB of context memory; with this platform, each GPU's accessible context memory skyrockets to 16TB. This poses a direct and significant benefit to the NAND market. Additionally, in the field of physical AI, NVIDIA has released an open-source model called "Alpamayo" for L4-level autonomous driving development, introducing chain-of-thought and vision-language action (VLA) models based on reasoning, attempting to give vehicles human-like thinking capabilities.

AMD: Lisa Su's race to catch up, targeting the 2027 showdown

AMD is desperately narrowing the gap, Lisa Su also views "agent AI" and "physical AI" as the keys to the next wave of growth. Its roadmap is exceptionally clear: the Helios rack equipped with the MI400 series GPUs will be launched in 2026, but this is just the appetizer. The real highlight is the MI500 series planned for release in 2027, which will be based on the next-generation CDNA-6 architecture, using a 2nm process node and HBM4E memory In the enterprise and PC sectors, AMD is trying to break through with cost-effectiveness. It has launched the MI440X GPU for enterprises and introduced the Ryzen AI Max for PCs (competing with NVIDIA's DGX Spark but at a lower cost).

Additionally, the Ryzen AI Halo development platform will be available in the second quarter of 2026, supporting local execution of models with up to 200 billion (200B) parameters, attempting to seize the initiative in the edge computing field.

Micron: DRAM is not just in short supply, but under a rationing system; NAND welcomes a new engine

If you want to find a textbook case of a "seller's market," just look at Micron. The supply/demand environment for DRAM is extremely strong, with prices holding firm. Micron expects a 20% growth in industry bit supply by 2026, but this is merely a drop in the bucket compared to the current unconstrained market demand, and the company has even entered a rationing supply mode.

More notably is the NAND market. While the market's focus was previously on HBM, driven by NVIDIA's release of a new type of contextual memory storage controller, the demand for SSDs in AI data centers will see a significant increase. This means NAND is no longer a bystander at the AI feast but has gained an additional growth engine beyond HBM.

Marvell: Aggressive M&A and expansion, connectivity is the core moat

Marvell is reinforcing its fortress in the data center connectivity field through capital means. The company announced the acquisition of XConn Technologies for $540 million, aiming to enhance its scale-up networking capabilities through XConn's PCIe and CXL switching technology. This deal is expected to start contributing revenue in the second half of 2026, reaching $100 million by 2027.

In terms of growth guidance, Marvell appears very aggressive. It reaffirmed its targets for the 2026/2027 calendar years (CY26/27): a year-on-year growth of 25%/40% for the data center business and 20%/100% for the custom computing business. Management specifically mentioned strong order momentum in December, demonstrating its dominance in the connectivity market.

Analog Chips: Silence after inventory bottoms out, the road to recovery is long

Unlike the frenzy in digital chips, the analog chip sector is in a suffocating "bottoming" state. Taking ADI (Analog Devices) as an example, its channel inventory is currently below 6 weeks, far below the historical normal level of 7-8 weeks. This is the first time ADI has seen channel sales to channels exceed channel sales out, indicating that inventory has become too low to meet current demand. However, even so, OEM customers have not yet begun to replenish inventory on their balance sheets, making the recovery appear unusually cautious.

ON Semiconductor expects a normal pricing environment in 2026, with low single-digit price declines, but legacy product revenues will face about $300 million in headwinds in 2026. Skyworks hopes for the resilience of its largest customer (referring to Apple) in the high-end market to withstand the demand destruction in the overall smartphone market

Synopsys: Deep Integration of Design and Physical Simulation

The battlefield of chip design is shifting.

Synopsys showcased the synergy achieved through integration with Ansys, with the first joint product targeting the advanced packaging field expected to be released in the first half of 2026. This marks a transition in chip design from purely logical design to the deep waters of integrated physical simulation; the future of EDA is not just about drawing, but also about precise simulation of the physical world.