Semicon West 2026: AI Compute, Jagged Edge LLMs, and Industry Shifts
Currently in Boston after a trip to San Francisco, the speaker shares insights and observations from recent events, including a keynote speech and attendance at Semicon West. This U.S. trip, occurring just weeks after a previous one, was primarily for a keynote address at the Test Vision Symposium at Semicon West.
Keynote at Test Vision Symposium
The keynote, titled "Shifting Left and Right, the evolution of ATE," traced the history of the automated test equipment (ATE) industry from its inception in the 1950s to the present day. It expanded on a previous video about the ATE industry, adding anecdotes from the 2000s onwards, particularly concerning test developments for mobile and AI industries. Although not recorded, there's a possibility it will be adapted into a future video. The speaker expressed immense honor for the invitation to deliver this talk.
Semicon West 2026 Impressions
The speaker arrived on October 14th, missing the first open floor day of Semicon West 2026, and felt that fatigue might have influenced their impressions. The overall sense was that the 2026 show marked "the quiet sunsetting of a different era."
Last year, there was a plan to alternate Semicon West between Arizona and San Francisco, but this plan seems to have been abandoned. Multiple attendees indicated that this would be the last San Francisco show. SEMICON West 2027 is scheduled for March 2027 in Arizona, a mere six months after the 2026 event. This quick turnaround is likely an attempt to find a less congested time slot, as SEMICON West 2026 overlapped with the OCP Global Summit in San Jose, which reportedly "stole its thunder."
The semiconductor and wafer manufacturing equipment industries are over 70 years old and well-established. In contrast, OCP addresses modern challenges like inference hardware, where essential vendors for areas such as cooling and networking are still being decided. The OCP show was sold out and described as "popping."
The move to Arizona also reflects a shift in the American semiconductor industry from its Intel-centric hub in Santa Clara to new production sites in the American hinterlands. The speaker anticipates the show will eventually move to Texas.
Shifting Moods: Inside vs. Outside Silicon Valley
Despite the short time since the last trip to Hot Chips, the mood has noticeably changed, revealing an intriguing split between Silicon Valley and the outside world.
Outside Silicon Valley: Unease and Financial Jitters
Outside the valley, things are "unsteady," with financial professionals feeling "antsy." The Financial Times reported Anthropic's numbers, which, while impressive (16 billion revenue with excellent gross margins), were deemed insufficient by the market, causing turmoil in AI trade stock prices.
The previous discussion about a potential compute glut in the second half of 2027 has now shifted to thinking about the supply chain for 2028. The goal of 50 gigawatts of compute by the end of 2027 is likely achievable with existing power infrastructure.
For 2028, incremental capacity estimates vary, but the target seems to be an additional 40 to 50 gigawatts. These numbers are still in flux, as final decisions for equipment production (which takes 1 to 1.5 years) won't be made until mid-to-early 2027. The question remains whether the industry will be convinced to add another 50 gigawatts on top of the 2027 capacity, or if the 2028 expectation is too low, implying flat or mild growth.
Semiconductor spending expectations are rising rapidly: - Lam Research (October 2026 earnings call): - 2026 wafer fabrication equipment spending: ~$150 billion - 2027 spending: >$200 billion - Goldman Sachs: - 2027 spending: $218 billion - 2028 spending: $281 billion - Whisper numbers from other firms: $300 billion or more
Despite these projections, there are "rumblings" of hiring freezes at wafer fabrication equipment companies, a "patchy signal" that suggests a bullwhip effect in the making, with companies bracing for potential demand fluctuations.
Inside Silicon Valley: The "Jagged Edge" and Compute Long
For builders within Silicon Valley, life continues, but the recurring question is "why are you doing this?" A common part of the answer is the concept of the "jagged edge."
The Jagged Edge
This concept describes how today's Large Language Models (LLMs) excel at certain tasks but perform poorly at others, often in non-intuitive ways. This "jagged edge" is why startups are encouraged to build their own specialized models, such as world models or models for handling specific cases. While frontier LLMs like Anthropic's or OpenAI's might be excellent at coding, solving complex math problems, or hacking websites, they supposedly "fall far short" outside their "golden path." The same term has been applied to Chinese open-source models, which show proficiency in some areas but significant deficiencies in others.
Going Long on Compute
Despite the compute glut concerns, there's a strong belief that not all projected gigawatts will materialize on time due to: - Profound power generation shortages - American labor issues - Data center NIMBYism (Not In My Backyard) - Shifting political winds
For these reasons, model training startups, including some of the "smartest people," are "going long on compute," believing that every incremental unit of compute will become even more valuable. The most sensible way to do this is by raising equity and minimizing debt, as missing debt payments can kill a company. However, some argue that not buying enough compute can also lead to failure by allowing competitors to outcompale on scale. This leads to the idea that simply raising $100 million for chips might not be enough; taking on debt to buy even more chips is necessary to be "truly long," otherwise one is "structurally short."
The Ecosystem of Subsidies
The speaker also explored how smaller companies building on top of frontier labs are operating, noting a significant dependence on subsidies provided by these labs through subscriptions.
Subscription Subsidies
Mid-2026 research by SemiAnalysis found that ChatGPT and Claude subscriptions offer significantly more value than API pricing. For example, a $200 plan could yield $2,000 to $14,000 in API value. Cash-strapped startups exploit this by loading up on multiple subscriptions, switching between them for development work, and "burning tokens especially hard when limit resets are announced."
The speaker acknowledges this as a startup strategy to achieve stability but expresses concern that cutting off these subsidies, suddenly or gradually, would severely impact these businesses. Alarmingly, some startups have even wired their customer-facing products to these subscriptions, running off subsidized tokens, which is deemed an "absolutely horrible and egregiously distorted way to run a business" and "unsustainable."
Parallels to Ride-Share Companies
This situation draws parallels to Uber and Lyft in 2015-2016, when they spent hundreds of millions on rider subsidies. Uber's leaked 2016 financials showed customers paying only 41% of trip costs, though this included extreme subsidies in mainland China. While AI subscriptions are likely not subsidized to that extent, price discrimination with higher-priced plans (like OpenAI's Pro 500 plan) is already emerging.
AI startups are preparing for this by basing product economics on API pricing and exploring orchestration strategies, such as combining smarter proprietary models with open-source or locally run sub-agents.
The Impact of AI on Mathematics
The impact of AI on the field of mathematics is widely recognized, with practitioners grappling with the possibility of their world being "thoroughly dynamited" in a short period. This mirrors the disruption experienced by chess and Go players when AI mastered their games, leading to a pivot where players became athletes and personalities. The speaker doubts this will happen for mathematicians ("Mathletes might become a niche thing"), but acknowledges stranger things have occurred.
The most likely outcome is a transformation of the math field itself, affecting different groups differently: - Older professors (e.g., Tao): Will likely be fine. - Younger PhDs: Will likely be fine due to time to adjust. - Those in the middle: Most affected, having committed to math as a career but not yet built a reputation.
The "Why" Behind Math Focus
The significant spending by OpenAI and Anthropic on math problems (e.g., Navier-Stokes costing an estimated $6.5 million in API token prices) raises questions about the economic impetus beyond mere marketing. While marketing and "love of the game" play a role (a 1% boost in evaluation could repay API costs), a deeper reason is suggested: Recursive Self-Improvement (RSI).
Recursive Self-Improvement (RSI)
RSI, distinct from repetitive strain injury or Relational Software Inc., is often imagined as an autonomously running loop constantly making improvements. While a "lights-out factory inside a data center" is the eventual goal, the current objective is more modest: getting LLMs to consistently prove hypotheses (like math proofs) true or false. Once achieved, LLMs could autonomously process human-generated hypotheses, proving them and feeding the learning back into the model.
Under this theory, RSI is essentially "automation for doing LLM training experiments." The speaker finds this idea compelling, especially since LLMs are fundamentally mathematical. This suggests that investment in solving math problems is a critical part of future R&D plans, rather than just marketing or "flexing."
A side theory posits that this work could be leveraged for formal verification of chip designs, though the speaker lacks sufficient knowledge to judge its validity.
RSI and Compute Utilization
If the RSI speculation holds true, it explains how Anthropic and OpenAI plan to utilize their "tsunami wave of compute" in the coming year: for RSI.
Previous assumptions of a 50/50 split between training and inference seemed problematic, as inference couldn't economically absorb that much compute. The solution to this dilemma is a revised split: closer to 75% training to 25% inference. All this training becomes feasible due to automated experimentation, i.e., RSI.
Investment Outlook and The "Jagged Edge" of the Economy
The speaker reiterates that this is not investment advice but notes a growing sentiment among professional investors that things are getting expensive. The AI trade is highlighted as one of the few "headlining economic trends in the US that still works." Outside of technology and AI, other sectors and countries present concerning "warts" and seem "broken."
However, even the AI trade is no longer as easy as it once was. Obvious plays like chips, power, optics, and memory have already seen their stock surges and now sit at elevated valuations. The question is, "What is next?"
In the private space, funding startups "going long on compute" is costly, and while appetite remains, the window for such funding might not stay open indefinitely.
Despite these challenges, the speaker remains optimistic, preferring it over being a "downer." However, the mood has shifted. "Gung Ho Valley still marches forward along the jagged edge, but falling off now somehow feels easier than before."
Takeaways
- The keynote at the Test Vision Symposium traced ATE history from the 1950s to modern AI and mobile test developments, highlighting the industry's evolution.
- Semicon West 2026 was seen as the “quiet sunsetting” of an era, with the show moving to Arizona for 2027 and likely later to Texas, reflecting a geographic shift away from Silicon Valley.
- Outside Silicon Valley, investors are uneasy about AI spending, with projected semiconductor equipment spending rising to over $200 billion in 2027 and concerns about a potential compute glut and hiring freezes.
- The “jagged edge” concept describes how current LLMs excel on a narrow set of tasks while failing elsewhere, prompting startups to build specialized models and to “go long on compute” despite power and labor constraints.
- AI startups heavily rely on subscription subsidies from services like ChatGPT and Claude, a model likened to ride‑share subsidies, which could become unsustainable if those discounts are reduced.
Frequently Asked Questions
What does the "jagged edge" refer to in LLM performance?
The "jagged edge" describes the uneven capability profile of large language models, where they excel on tasks within their trained “golden path” such as coding or math, but perform poorly or unpredictably on many other tasks, leading startups to create specialized models.
Why is Semicon West moving from San Francisco to Arizona and possibly Texas?
The move reflects a strategic shift to avoid scheduling conflicts and congestion, as the 2026 show overlapped with the OCP Global Summit and faced “sunsetting” of the San Francisco era; relocating to Arizona (and later Texas) offers a less crowded slot and aligns with the broader migration of U.S. semiconductor production away from the Intel‑centric Santa Clara hub.
Does this page include the full transcript of the video?
Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.
remains whether the industry will be convinced to add another 50 gigawatts on top of the 2027 capacity, or if the 2028 expectation is too low, implying flat or mild growth. Semiconductor spending expectations are rising rapidly: - **Lam Research (October 2026 earnings call):** - 2026 wafer fabrication equipment spending: ~$150 billion - 2027 spending: >$200 billion - **Goldman Sachs:** - 2027 spending: $218 billion - 2028 spending: $281 billion - **Whisper numbers from other firms:** $300 billion or more Despite these projections, there are "rumblings" of hiring freezes at wafer fabrication equipment companies,
"patchy signal" that suggests a bullwhip effect in the making, with companies bracing for potential demand fluctuations.
is "why are you doing this?"
common part of the answer is the concept of the "jagged edge."
Helpful resources related to this video
If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.
Links may be affiliate links. We only include resources that are genuinely relevant to the topic.