bubble boi
Semiconductor Market Commentary, Thru Project Mascot
Astera Labs' CXL integration thesis faces a threat if Nvidia builds a CXL-attached memory architecture: Jensen would not use Astera for it. Such an architecture could make training and data sharing across nodes easier. Nvidia's Feynman might also get a better I/O die to avoid switch hops, as Google did with TPUs. x.com x.comÂ
Astera Labs is the only good CXL play because it is building the whole scale-up stack, not just selling the protocol. CXL rides the PCIe physical layer and its IP is sold by Cadence, Synopsys and Rambus, so IP alone has no moat. Astera's CXL switches, Scorpio fabric and COSMOS software make networked CXL cards usable and sticky; they are not major revenue yet but will be. Revenue hit $308M last quarter, up 93% YoY, into a scale-up switch market projected at $20B by 2030. x.comÂ
Astera Labs carries a huge premium for flawless execution against a CXL market that is small near term. Switch-based CXL memory pooling capturing $15-16B in 2026-2028 is low probability: Marvell guides its XConn switch business to only about $100M by fiscal 2028, PCIe included. Pooling software likely stays immature into 2028, per a Samsung white paper, and Nvidia is contesting the warm KV-cache tier. Astera does CXL best, but SK hynix and Micron offer better risk-adjusted returns. x.com x.comÂ
Intel and Rambus could explode if Nvidia announces a co-designed Rubin Ultra or Feynman product with Intel, which looks like a matter of waiting. Agentic AI demand is through the roof, and Nvidia's Vera CPU volumes should exceed its VR200/300 volumes at maturity. x.comÂ
Inference's move toward edge devices and dense, memory-hungry racks should drive demand for specialized chips, links, power delivery and cooling. McKinsey projects inference at 60% of AI computing demand by 2030. Ambarella makes on-device AI chips; Astera Labs links GPUs, memory and networks; Credo supplies short copper connections; Rambus earns from memory interfaces; Navitas reduces power waste; Modine cools dense racks; and Super Micro builds inference-specific server racks at volume. x.com … Show more
Qualcomm is building the network layer around its HBC memory architecture as an integrated system, attacking AI data movement at several layers at once. HBC cuts movement between memory and compute, Alphawave adds SerDes, optical and die-to-die links, and per AIwire Cornelis Networks will provide networking for Qualcomm's Dragonfly HBC accelerators, with compute in the fabric to cut stalls. Qualcomm's data-center chief Tony Pialis says moving data across the rack is now as important as compute. … Show more
Nvidia's case is a lot weaker than two years ago, because large providers such as OpenAI and Meta are likely to design custom chips tuned to their own models. Nvidia wins only if workloads and models stay diverse. The claim that frontier labs are decoupling from Nvidia is doubtful, since OpenAI's recent efficiency gains likely came from Blackwell, but it holds for Google and Amazon. Nvidia also needs to own the open-source ecosystem, which is contradictory by definition. x.com x.comÂ
AI's biggest bottleneck is power and cooling, not memory, making Vertiv, GE Vernova, Bloom Energy and Eaton buys. Musk expects a prolonged compute shortage; Nvidia, Broadcom, Micron, SK hynix, TSMC, Arista, AMD, Marvell, Credo, Lumentum, ASML and Astera Labs could benefit massively. Grok Bot's 800,000 paid weekly users have more than doubled in a month; Musk says agents could eventually use more compute than any other part of SpaceX's Super Intelligence business, benefiting Dell and Sandisk too. … Show more
Nvidia remains the core of the AI supply chain: in Morgan Stanley's bill-of-materials teardown of the next-generation Rubin VR200 rack, GPUs are still about half the total cost, as Nvidia sells rack-scale AI infrastructure. The teardown also shows HBM with the largest cost increase, NVLink Switch and other networking chips roughly doubling, and PCB content up 233%, so value in rack-level upgrades is spreading to memory and networking as bottlenecks shift to communication between GPUs. x.comÂ
Intel and AMD face a real shift in competition as NVIDIA moves into CPUs faster and more directly than expected. On NVIDIA's earnings call, Jensen Huang said its new Vera CPU platform will contribute $20 billion of standalone revenue this year. The pressure on Intel and AMD is not just fund rotation but a genuine change in the competitive landscape. x.comÂ
Astera Labs and Montage Technology are positioned to benefit as trillion-parameter open-source models push AI inference bottlenecks from compute to memory. Open models grew from 1T to 2.8T parameters in under a year and now rival frontier models. In MoE models over 98% of expert weights are cold at any moment, so keeping hot experts in HBM and warmer ones in CXL-attached DDR5 can cut costs sharply. Astera Labs leads CXL controllers for global hyperscalers, Montage for China's supply chain. x.comÂ
Intel's turnaround has validation the market has not fully priced in, as Nvidia is already trialing Intel's advanced packaging and 18A process for next-gen Feynman GPUs launching in 2028. Google's order for more than 3 million TPUs in 2028 made for another big day for Intel. The same setup preceded generational runs in AMD, Nvidia, Micron and SanDisk. x.comÂ
