Hot Chips 2026 & Semicon Taiwan: AI Chip Startups and Compute Shortage

This article summarizes observations and conversations from recent trips to Hot Chips 2026 in the Bay Area and Semicon Taiwan, highlighting key trends and developments in the AI and semiconductor industries.

Hot Chips 2026: A Packed House and Emerging Themes

The Hot Chips conference has grown significantly. The auditorium, once sparsely filled, is now packed, indicating a surge in interest and attendance, including a notable increase in finance professionals.

Key themes from the conference included:

• AI in EDA (Electronic Design Automation): The impending integration of AI into EDA processes was a major topic.
• Silicon Photonics: While not as prominently featured as the previous year, silicon photonics continued to be a significant area of discussion, with many experts present.
• Memory Sessions: Presentations by Micron, Samsung, and SK Hynix on memory technology were exceptionally well-attended, reflecting the current boom in the memory sector.

OpenAI's Jalapeno and the Rise of New AI Chip Startups

A highly anticipated event was OpenAI's presentation of its new AI chip, "Jalapeno." This presentation, though brief, generated considerable interest. The author, who previously believed the window for new AI chips had closed (with the exception of Google's TPU), now retracts that statement, acknowledging the current hot market for AI-focused chips.

Beyond OpenAI, a new cohort of AI chip startups is emerging, including Etched, Matt X, Posetron, and Fractile. Etched, in particular, made an impression despite not having a formal presentation, with its representative actively participating in Q&A sessions. These companies are not just designing chips but also building the entire rack and system infrastructure, a complex undertaking that requires significant talent and execution. The challenge of acquiring top talent in this space is frequently discussed.

The survival of these new companies hinges on execution and their ability to navigate a constrained supply environment, particularly concerning access to High Bandwidth Memory (HBM) and successful tape-outs.

Semicon Taiwan: Mayhem and TSMC's Influence

Semicon Taiwan has transformed from a "humble affair" in 2022 to a "mayhem" of activity. The event now fills multiple exhibition halls, with heightened hype and elaborate displays, including "showgirls" and attendees in bunny suits. A "dream fab" concept, allowing visitors to experience a simulated fab environment, was completely booked.

Notably, major players like TSMC and ASML do not have traditional booths. Despite this, TSMC dominates the show. Their teams actively scout new vendors and technologies, and their executives headline many information sessions, openly outlining their challenges to the industry. This suggests that Semicon Taiwan effectively serves as TSMC's "request for proposal" to the entire semiconductor ecosystem.

The Future of Semiconductor Manufacturing: Bigger Chips and Advanced Packaging

The overarching technological trend is towards larger chips. AI chips are pushing the limits of existing lithography machines, particularly the reticle limits. Emerging technologies are focused on either overcoming these limits or scaling the ecosystem to accommodate them, with advanced packaging being a critical area.

TSMC has indicated plans for "battleship-sized chips" spanning up to 14 reticles. The industry is rapidly responding, with developments like panel-level packaging, a theme from Semicon Taiwan 2024. While interposers traditionally use silicon, the demand from AI systems is driving a shift towards larger, square-shaped glass panel interposers. The ecosystem is already adapting, with handler robots and other equipment designed for square panel sizes. This "chip-maxing" trend is seen as one of the most significant technological transitions in the semiconductor industry since the adoption of 300mm wafers.

The Bay Area: AI Agents, Compute Shortage, and Neoclouds

The Bay Area has seen a significant shift towards long-running AI agents, spurred by the release of more capable models like Anthropic's Opus 4.5 and the "Open Claw" development. This has led to a scramble for compute resources, with prices rising and the spot market for compute vanishing.

"Neoclouds," a term coined by SemiAnalysis's Dan Nystedt, refers to pure-play cloud providers that resell AI training and inference compute (e.g., Coreweave, Nebius). These firms are confident that the compute shortage will persist for at least one to two years, leading many to attempt to start their own Neocloud ventures.

For AI startups, the compute shortage poses a significant business risk, potentially delaying product launches. Venture capitalists are now actively involved in securing compute for their portfolio companies, even resorting to unconventional methods like sourcing GPUs from unexpected locations.

In contrast, major AI players like Anthropic and OpenAI are "feasting" on compute, with an estimated 5 GW each. However, anecdotal evidence suggests that this abundance of resources may not always be used with optimal efficiency.

Google's AI Frontier and the Rise of SpaceX's AI

A common question posed to AI professionals was whether Google would return to the AI frontier. The general consensus was negative, with explanations pointing to Google's decision to sell compute to Anthropic as a sign of their own lack of belief in their frontier capabilities. Factors cited include substantial talent loss and a culture not conducive to frontier AI development.

If Google is not returning to the frontier, the question then becomes who in the US (excluding mainland China) is most likely to join Anthropic and OpenAI. The consensus points to SpaceX's AI efforts. Elon Musk's motivation and significant investment in acquiring talent (e.g., Cursor) and compute suggest a strong push in this area, evidenced by products like Grokbot.

Neolabs: Asymmetric Bets on AI Research

A fascinating development is the emergence of "Neolabs" – pre-revenue research startups funded to explore intriguing AI ideas. These are often characterized by elite researcher founders and a minimal business plan beyond discovering new structures. Examples include Safe Superintelligence (which raised a billion dollars without releasing a product), Recursive Intelligence, Humans, Core Automation, and Flapping Airplanes. Other AI labs with products include Sakana AI, Thinking Machines, and Reflection AI, as well as domain-specific labs in material science, world models, and robotics.

While some might view this asset class as speculative due to the lack of immediate revenue, the rationale is that traditional sources of groundbreaking scientific research (universities, government labs, Bell Labs) have either faced funding cuts or talent drain. In the current AI scaling era, research is often proprietary. Neolabs offer a way for entities like Nvidia (with compute and cash) and VCs to fund asymmetric bets. OpenAI and Anthropic, being locked into existing paradigms, are less likely to explore these radical ideas. The funding, often in the form of compute, allows smart individuals to pursue potentially transformative discoveries. If a Neolab uncovers a compelling breakthrough, the giants are likely to acquire them at a high price, justifying the initial investment. This approach is seen as a better use of capital than mere share buybacks.

ROI of AI and the Compute Glut Prediction

A persistent question is whether customers are seeing a return on investment (ROI) from their AI token consumption. The answer remains "nebulous," with no clear metrics and reliance on anecdotes. While AI is lowering barriers to coding and impacting knowledge work (e.g., financial reporting), its impact on other areas like spreadsheet or presentation creation is less noticeable. OpenAI's internal use of agents has accelerated research, but their capabilities are still likened to an intern. The lack of "human research taste" in AI is noted, though some believe it can be achieved with sufficient scale.

Despite the perceived efficiency of AI, many in San Francisco report having more work than ever, suggesting that AI might be creating more tasks rather than automating them away.

The author challenges the widespread belief in a prolonged compute shortage. While a shortage exists now, the supply chain is responding aggressively. Estimated usable AI data center capacity is projected to surge from 15 GW at the end of 2026 to 45-55 GW by the end of 2027, representing a 30-40 GW increase. Even if OpenAI and Anthropic consume a significant portion of this (e.g., 30 GW), the projected revenue figures (e.g., $750 billion to $1.5 trillion for the two companies in 2027) seem implausibly high, straining credibility. The author believes the revenue per gigawatt will be significantly lower than current models suggest.

From an infrastructure perspective, 40 GW of new capacity in 2027 would require $800 billion in payments to infrastructure providers. This figure also seems unsustainable, implying that rental prices for compute must fall, leading to potential losses for some investors. The author predicts that the massive influx of compute in 2027, particularly in the second half, will shift the market from a shortage to a glut. While this doesn't necessarily mean the AI bubble will pop, it suggests an oversupply of compute. However, a subsequent phase of AI growth, perhaps driven by "persistent AI," could reignite demand.

Anti-AI Sentiment: A Growing Concern

A significant and concerning trend is the growing anti-AI sentiment. This negativity, exemplified by accusations of plagiarism and data theft surrounding a recent OpenAI proof for the Navier-Stokes problem, is becoming more prevalent online and among the general public. Reasons for this sentiment vary, including dislike of data centers, opposition to billionaires, concerns about AI "stealing" human knowledge, or a belief that AI is unnecessary. Regardless of the underlying causes, this sentiment is growing and has the potential to influence political outcomes. The author warns that those within the "SF bubble" should pay closer attention to these external feelings, as they can "metastasize with scary speed."