Tony Wang

I don't get enough credit for not stepping on my dog.

@jhong What’s a good historical example of this? The best I can think of is housing.

The demand for AI is going to grow exponentially for a while (if not forever). The supply for AI is going to look a lot more linear in comparison. Scaling atoms is a lot harder than scaling bits. For this reason, the price of compute is going to rise substantially. A lot of electronics will get expensive, and fun toys like image and video generation that are relatively cheap today may become prohibitively expensive. Enjoy it now while you can.

Tesla prototype (for robotaxi?) spotted in the wild. Notice no side view mirrors, two seater, smaller than model 3.

Jensen Huang at GTC2026 says he sees $1T in demand for Blackwell and Rubin by 2027: “In fact, that is going to be short.” (50% of their business are top 5 hyperscalers) $NVDA

@EthanChoi7 Good write up Ethan. I also think it’s important for higher education to keep up and look forward to seeing more great examples.

After this X post hit a nerve, I finally posted my article to answer one of the questions I get the most - "How do I AI-proof myself or my kids?". Hope it's helpful...

@Noahpinion They don’t let the best teas out of Taiwan. The domestic demand outstrips supply regularly.

This is the best tea I've had in my entire life, and it's actually pretty cheap. redblossomtea.com/products/formo…

@grok Are you sure that’s right? Can you check and confirm?

@TonyW Those 250k shares from '93, after NVIDIA's full split history (480:1 cumulative multiplier pre-2000), equal 120 million shares today. At current ~$185 price, that's ~$22.2 billion.

Jensen on stage with Synopsys CEO and says NVIDIA gave Synopsys 250k shares of NVIDIA when NVDIA was worth $10M but now Synopsys can’t find the shares. 🤯

@grok How much would that be worth today at today’s $NVDA price?

Yes, Jensen shared this exact anecdote publicly at SNUG'24 on March 20, 2024, during a talk with Synopsys execs. He said he gave them 250k shares in 1993 after their van showed up with tools, and that Aart de Geus had been hunting for the certificates in his desk drawers weeks earlier. Covered right after in EDA sites like DeepChip and SemiWiki. Not mainstream, so many $SNPS folks may be learning it fresh today. Bulls will see it as proof of unbreakable early ties (now supercharged by the $2B NVIDIA stake); bears likely won't care as it's 30+ year old history. No real price reaction expected.

@Tesla_Teslaway Great overview, where can I find the deep dives?

@rswillis96 @Tesla_Teslaway This is interesting. I’ve got a materials folks you should meet.

A4N’s Position Across the Bottleneck Stack Bottlenecks 2-3 (Memory Wall & HBM Supply Shortage) — This is the most direct hit. Every HBM stack is a vertical tower of DRAM dies bonded together, and each layer generates heat that compounds upward. The thermal filler materials between those dies need to be ultra-high-purity alumina to conduct heat without introducing electrical interference. The thesis you’ve been building around HBM packaging wastage as a hidden demand multiplier sits right here — as HBM goes from 8-hi to 12-hi to 16-hi stacking, thermal management per stack becomes exponentially harder, not linearly. More layers = more thermal filler material per unit, and tighter purity specs. Bottleneck 3-1 (PIM & Hybrid Bonding) — This is where it gets really interesting for the master thesis. Hybrid bonding eliminates traditional bumps and puts chips in direct contact with hundreds of thousands of I/O points. The thermal density at those interfaces is extreme. The material filling the gaps between those bonded layers needs to be: ∙Thermally conductive ∙Electrically insulating ∙Ultra-high purity (4N+ / 5N) That’s A4N’s exact product specification sweet spot. PIM (processing-in-memory) adds compute inside the memory stack, which means even more heat generation within the HBM tower itself. Every architectural innovation described in 3-1 makes the thermal filler TAM larger. Bottleneck 4 (Advanced Packaging / CoWoS) — The silicon interposer that sits beneath the GPU and HBM stacks on a CoWoS package needs thermal interface materials. As these packages get larger (reticle-limit and beyond with chiplet stitching), thermal management across the entire substrate becomes a system-level problem. HPA-based thermal fillers and CMP slurries for planarizing those interposer surfaces are both A4N addressable markets. Bottleneck 5 (Power Wall) — A 1000W GPU in a rack of 72 (think GB200 NVL72) is generating ~72kW per rack. The cooling hierarchy runs: die → thermal filler → package → heatsink → liquid cooling → facility. A4N’s materials sit at the very first thermal interface — the most critical one. If the thermal filler fails or underperforms, no amount of liquid cooling at the facility level saves you. Bottleneck 7 (Chiplets & CFET) — As the industry shifts from monolithic dies to chiplet architectures, you multiply the number of thermal interfaces. A single monolithic GPU has one die-to-substrate interface. A chiplet-based design might have 4-8+ dies on a single package, each needing thermal management. More chiplets = more HPA demand per package. The Key Insight for the Master Thesis What this timeline reveals is that A4N isn’t a single-bottleneck play — it’s a persistent, compounding demand story across bottlenecks 2 through 7. Every solution to the current bottleneck (more HBM stacking, hybrid bonding, bigger packages, chiplets) creates more thermal interfaces requiring ultra-high-purity alumina, not fewer. The semiconductor industry is solving bandwidth by stacking vertically and packaging more densely. Both strategies concentrate heat. The thermal filler TAM grows with every architectural innovation on this timeline.

@dickc Did you forget to log into @fakesacca before tweeting this

@hardmaru @Citi Let’s go! What a privilege to see you and the team continue to build.

Sakana AI started as a pure AI R&D lab in Tokyo. Since then, we have built a solid enterprise business in Japan, working with major institutions like MUFG, SMBC, and Daiwa Securities. Now, we are ready to deploy our technologies abroad with @Citi!

@RealJGBanks Wild that $NVDA is not on here

BREAKING: Wall Street thinks the pipes behind AI are entering a multi-year run. Here is the Full Cheat Sheet: THE CABLES (Optics Upgrades) $LITE - Lumentum
$COHR - Coherent
$AAOI - Applied Optoelectronics
$FN - Fabrinet
$CIEN - Ciena
$INFN - Infinera THE FIBER (Physical Backbone) $GLW - Corning THE NETWORK (Moves the Data) $ANET - Arista
$AVGO - Broadcom
$MRVL - Marvell
$CSCO - Cisco THE SERVERS (Houses AI) $SMCI - Super Micro
$DELL - Dell
$HPE - Hewlett Packard Enterprise THE CHIP BUILDERS (Make It Possible) $TSM - Taiwan Semiconductor
$AMAT - Applied Materials
$LRCX - Lam Research
$KLAC - KLA THE SPENDERS (Who Pays for It) $MSFT - Microsoft
$AMZN - Amazon
$GOOGL - Alphabet
$META - Meta

@credistick Indeed. Morris Chang has often credited HSR as one of the factors contributing to the success of TSMC

@TonyW I love that there is research out there to answer questions like “what’s the influence of high speed rail on VC?”

Will distributed VC partnerships generate more alpha in a future that is increasingly well connected and polycentric? Venture capital is one of the most geographically concentrated strategies in finance. More than half of the VC offices in the US are in just three cities: San Francisco, Boston, and New York. California alone accounts for over 50% of VC firms and 64% of all investment dollars. Add New York and Massachusetts and three states contain over 70% of all firms deploying 80% of all capital. However, that concentration might be fraying as California's proposed asset seizure tax has triggered capital dispersion. The beneficiaries are peripheral US hubs like Austin and Miami. Meanwhile, US investors are increasingly looking across the Atlantic to hubs like London, Munich, and Zurich that offer deep pools of technical talent (particularly in AI/HALO categories) with more reasonably priced equity. Indeed, a growing body of research, stretching from the United States to Europe and China, indicates that the relationship between proximity and performance is misaligned with today's capital concentration. 1) Non-local deals outperform The most important finding comes from a study of 28,434 venture capital investments across 2,039 firms. VC firms in the major venture hubs do tend to be high performers, with a success rate 4.4 percentage points higher than elsewhere, but the data reveals something counterintuitive about where that edge comes from. Investments made in a firm's main office region have a success rate of just 14.5%, while branch office and outside investments both succeed at roughly 17%. "Surprisingly, much of the VC outperformance in these venture capital centers arises from their non-local investments. This finding is counterintuitive, since venture capitalists might be expected to be the most involved and add the most value to the geographically closest companies." (source: "Buy Local? The Geography of Successful and Unsuccessful Venture Capital Expansion") The pattern holds across stages. On early-stage deals, firms in major hubs succeed at 15.1% versus 11.3% for others. On late-stage deals, 20.7% versus 15.7%. In multivariate analysis, when the authors interact location with investment type, the outperformance is almost entirely explained by non-local deals. The proposed explanation for this is investor discipline; distant investments carry higher monitoring costs, so VCs demand stronger prospects. "The higher rates of return on non-local deals may indicate economically meaningful geographic differences in the availability of venture capital." As an example of this, PitchBook analysis in 2018 ranked Chicago at #8 in the US for VC investment, but #1 for VC returns (exit value reflected as MOIC). The top hubs for deployment do not produce the most efficient returns. (source: “The Bay Area & beyond: Ranking US metro areas by VC invested and returns”) 2) Syndicates are designed to enhance reach A study by Dolencic across 11,017 U.S. venture capital investments finds three variables that significantly increase the distance between investors and portfolio companies. On the national level, and for out-of-state investments from California and Massachusetts, experience has a positive effect on distance, as seasoned VCs are less intimidated by information asymmetry. Syndicate size is consistently positive across all specifications. More co-investors imply greater geographic diversity, and a distributed partnership may represent a more appealing syndicate partner due to their enhanced reach. "Syndication is important in this context as it may help to overcome distance. Sorenson and Stuart suggest that a venture capital firm that is a part of a large syndication network can find potential interesting investment opportunities located farther away. Second, syndication may help overcome the agency costs related to the greater distance, because a syndicated partner that is located closer can help with monitoring and oversight." (source: "Factors Influencing the Geographic Scope of Venture Capital Investments") The conventional wisdom, captured by Zook's interview with a San Francisco founder, holds that proximity is non-negotiable ("I know some venture firms that say, 'if I cannot drive there within an hour, I don't make the investment'.") but the data shows that experienced, well-networked firms routinely overcome this constraint. 3) In polycentric markets, connectivity > proximity Another challenge to the proximity orthodoxy comes from a survey of 85 investment managers in Germany. Independent VC firms keep less than 30% of their portfolios within 100 km, with the rest spread across Germany and abroad. Banks and business angels, by contrast, keep more than 75% of investments within that radius. "The analyses clearly showed that the importance of regional proximity between the VC firm and its portfolio companies is widely overestimated in the literature. The VC companies in Germany, especially the private and independent VC firms, do not focus mainly on investments located nearby." (source: "Does Venture Capital Investment Really Require Spatial Proximity?") As one manager wrote on his questionnaire: "It is not time to pick and choose in the regional sense as long as you want to earn money." The primary challenge of managers in polycentric markets is the wide distribution of good opportunities, which is important to consider as the US heads in that direction. Building on this, a 2025 study by Liu, Chen, and Cao on 402 Chinese firms listed on the SSE STAR Market finds that high-speed rail connections between VC firms and portfolio companies are associated with better outcomes via more effective knowledge transfer. "Examining the relationship between VC firms and investee high-tech enterprises from the perspective of knowledge transfer, we argue that the time-space compression effect of HSR connections between VC firms and investee enterprises during the postinvestment stage facilitates greater knowledge transfer from the former to these enterprises, thus improving their innovation performance" (source: "Geographic network characteristics of venture capital firms and high-tech enterprise innovation") The effect is strongest for early-stage companies in emerging industries. It turns out that what matters is the ease and frequency of connection, not a need for permanent co-location. How did we get here? While the studies above treat geography as static, additional research suggests that the influence is cyclical, with benefits of concentration amplified in hot markets and reversed in cooler periods. A 2017 study found that network centrality, a property closely correlated with geographic concentration in hubs like Silicon Valley, drives superior VC performance, but only for specific periods. Centrality advantages hold robustly in boom years (the late 1980s, the 1995-2001 dot-com era) when interest rates were low, capital was abundant, and concentrated interest in hot sectors rewarded the dense, fast-moving networks that hub-based firms excel at. In bust periods (1991-1994, 2002-2008, 2022+), those same centrality advantages may turn negative. The point estimates on centrality's effect on performance actually flip sign during downturns. “When we examined a longer range of data and in shorter subsets, centrality often had no relationship—and sometimes even had a negative one—with firm performance. The positive relationship appeared most strongly during the boom period of the late 1990s.” (source: “The Stability of Centrality-Based Advantage”) The implication is that the concentration-performance link documented in some literature may be an artifact of the market conditions in which it was measured. The period from 2009 to 2021, the zero-interest-rate era that shaped the modern VC industry, was the longest and most capital-rich boom in venture history. Concentration thrived because velocity was a competitive advantage. Funds clustered in the same hubs, chased the same sectors, and used capital itself as a weapon to inflate valuations. In that environment, being embedded in a dense local network, where information about the next hot round moved at the speed of a group chat, was worth more than being distributed across markets with differentiated deal flow. However, that environment no longer exists. Today, the structural advantages of distribution, access to undervalued opportunities in peripheral hubs, diversification across sectors and geographies, and the higher hurdle rate that distance naturally imposes, become more valuable relative to the coordination speed that concentration provides. If the boom-era premium on concentration was partly about moving fast in crowded markets, the post-boom premium may belong to firms that see more of the landscape and pick more carefully. What does this mean for firm design? Taken together, these studies describe an industry where concentration has persisted out of habit and network effects rather than performance. The best returns come from non-local deals, while experience, connectivity and effective syndication alleviate the frictions created by distance. The real constraint is the ability to find high-quality opportunities, wherever they may be. So, hypothetically, a small firm that was distributed across primary and secondary hubs could generate impressive alpha. e.g. One partner in San Francisco or New York for LP access and deal density. Another in London or Munich for European deal flow. A third in a peripheral hub, like Austin, Stockholm or Boston. Each partner provides local-grade sourcing and founder relationships in their region while applying the discipline and selectivity of a distant investor in their colleagues' deals. There are operational frictions (coordination costs, cross-border complexity, time zones) to overcome, but these are not structural disadvantages. The structural advantages, however, (wider deal flow, higher selectivity per geography, multiple network positions, and the natural quality filter of distance) are hard for a geographically concentrated firm to replicate. In an increasingly distributed world, where founders build from Berlin and Pittsburgh as readily as Palo Alto, a venture firm that is more geographically mobile might hold an asymmetric advantage. Research suggests this advantage is meaningful, measurable, and currently underexploited.

@Trace_Cohen How large of a fund is this? This doesn’t sound like a typical seed fund, but just realized that why you posted it lol

Just saw a Seed VC lead a seed round with $5M and also add in a $5M uncapped SAFE to ensure that they can get their basically full pro-rata at the next round. Otherwise when the investment works and raises more, they most likely couldn’t invest that much.

@jinen Just DMd

@TonyW TONY! thank you, and yes please! it's been way too long!

big news: we're spinning out gather into a profitable independent company, and our AI team is joining figma! building gather has been the honor of a lifetime. i'm deeply grateful to the team and customers who made it all possible. i’ll be a townsperson forever, but i’m excited to become a figmate soon 🫡. more in our founders' letter below.

@chris_j_paxton @lukas_m_ziegler This is the right answer

@lukas_m_ziegler nvidia

if you could bet on just ONE company in robotics - what would it be?