reSee.it - Related Video Feed

The conversation centers on how quickly Chinese open AI models are advancing and whether China will reach or surpass AI leadership. Eric Schmidt is cited as making several timeline corrections after earlier claims that America was about five years ahead of Google’s AI relative to China; the gap was later revised from about a year to months and then to weeks. The discussion also references the release of models such as Babel and GLM 5.2 “neck and neck,” raising the question of whether a crossover point will occur and whether China will take the AI lead afterward. A key factor discussed is the AI inference hardware supply chain. Previously, NVIDIA was described as the dominant single supplier whose hardware ran AI inference. The speakers say other manufacturers are now figuring out how to make chips that aren’t NVIDIA, which would break a single-hardware bottleneck and shift toward a “plethora of chips” competing through an open market rather than a centralized hardware cartel. AMD is then discussed as a strong player in hardware for AI-related workloads. One speaker says AMD’s CEO “looks kind of like Jensen Huang” because they are described as cousins from the same Taiwan family, competing on different hardware branches. The focus is on AMD’s development of high-bandwidth, unified RAM and large memory capacity, including a 192 GB unified platform mentioned for “Strix Halo,” positioned as fast for personal use rather than replacing data centers. The speakers contrast hype claims that consumer hardware can fully substitute for data centers with the idea that it can still be useful. On local AI performance, the discussion turns to token throughput. One speaker argues that with limited token rates, a powerful model can run on a “very powerful Macintosh,” but for real work they want roughly 100 tokens a second or 200 tokens a second. Another speaker notes that most people operate around 25 tokens a second. The conversation then describes “agent swarms” that run multiple steps: agents inspect codebases, find bugs, apply fixes, perform code review, and finalize changes. This pipeline, they say, would not run locally at 26 tokens a second; instead, it would take about a week rather than an hour. The speaker cites OpenAI token usage, stating someone put in “a billion tokens last week,” and compares this to the 26 tokens per second constraint, concluding that the computation would take an extremely long time.

Speaker 0 said they downloaded all the open weights of GLM 5.2 and intend to run it one day, noting they currently lack hardware. They also argued that the U.S. banning Anthropic models effectively hands AI implementation’s future to China. Speaker 1 referenced a Reuters story on June 17 stating the Trump administration decided not to ban DeepSeek R1 “yet,” implying a ban may come later, similar to actions taken with TikTok. They said models such as Anthropic’s “fable opus 4.8” are very expensive, while Chinese models including DeepSeek R1, GLM, Qwen, and Minimax M3 are available at a fraction of the price. They argued these Chinese models have improved and are now almost at the level of what the U.S. frontier can produce. Speaker 0 agreed on the cost gap, stating that in some cases it is “50 times less” and sometimes even higher. Speaker 1 then questioned why companies pay more via employee salaries or token usage when Chinese models can perform similar tasks for much less. Speaker 1 cited an Nvidia CEO claim that if a $500,000 employee is not spending $250,000 in tokens, they need to be fired, adding that $250,000 exceeds what many engineers make as salary. They argued that if similar performance can be achieved far cheaper, spending at the higher level becomes harder to justify. Speaker 1 proposed the U.S. will respond with an import ban and controls akin to “the Great Wall” and “the Great Firewall of America.” They said it would begin with a blacklist blocking access to certain services or websites, progress to whitelists allowing access only to government-approved entities, and then declare open models “problematic and unsafe.” They said the U.S. would require entities to prove open models can “naturally run within the borders of the United States,” and if not, would remove them from open-source repositories such as Hugging Face. Speaker 0 challenged whether this would include stripping models from Hugging Face, controlling GitHub, and criminalizing downloading open weights from China; Speaker 1 replied that this is exactly what they believe would happen in stages. Speaker 1 argued that it would be difficult to determine what models do when only the model weights are available, describing models as “black boxes” and noting concerns about malicious intent embedded in weights. They added that even OpenAI and Google do not fully know what their models are capable of and said static analysis for model forensics is an unresolved “frontier question.” They concluded that Chinese companies or the Chinese government proving open models are harmless and contain no malicious intent is “virtually impossible” given this problem.

The speaker says they “buy the fact” that SpaceX is a solid company with a great business plan that will do extremely well, and that they leave the price to the market. They add two quick points about what SpaceX is. First, when people ask “what is SpaceX?” the speaker notes it’s often described as a rocket company that will take astronauts back to the moon and as having great partnerships with NASA. They argue that it is “so much more than that,” emphasizing that Elon Musk is putting data centers into space and using SpaceX rockets for that purpose. The speaker frames the key advantage as “unlimited free power” from solar power in space, where conditions are “freezing cold,” reducing the need to spend money or energy heating or cooling systems. They assert that, in space, constraints faced by massive data centers on land do not apply in the same way. Second, the speaker explains that massive data centers on land face constraints including water, energy, chips, cooling systems, and local resistance from citizens. They highlight that power input and the energy source are major issues, and that water for cooling is particularly scarce. They state that these problems are not present to the same extent in space. They conclude that while SpaceX is a rocket company, it “might be the world’s biggest data center company.”

Speaker 1 explains that when he says the Earth’s magnetic field has remained roughly constant over long timescales, he means its magnitude is roughly constant on those scales, though it varies and undergoes reversals where the North and South Poles flip. He notes that reversals correlate with ice ages and other climate signals, but averaging over these fluctuations keeps the amplitude roughly constant. He emphasizes that without a dynamo, the field would diffuse away in about 10^5 years, leaving Earth unprotected from cosmic radiation, which would be harmful to life. Speaker 3 asks about the use of quantum computing in plasma physics, acknowledging its newness. Speaker 1 answers: We can’t use it right now. The short answer is “we cannot.” The longer answer is that it may take twenty years for a quantum computer to become useful for solving real problems. It would be a mistake to wait twenty years and then try to port existing codes to a quantum computer, because quantum computing has a fundamentally different architecture. Therefore, two lines of thought should develop in parallel: by the time a useful quantum computer exists, we should already know how to map our problems to it. Speaker 1 elaborates that solving nonlinear problems on a quantum computer is not straightforward. He discusses the challenge of devising quantum algorithms for nonlinear problems. He mentions working with the Madelung transformation, which maps the Schrödinger equation into fluid-like equations, noting that this approach is interesting because magnetohydrodynamics (MHD) equations are similar in some ways. While the Madelung transformation has limitations, it illustrates the kind of problem mappings that might make certain problems more tractable on a quantum computer, though this represents a completely different paradigm from conventional computing. Speaker 3 thanks Speaker 1. Speaker 2 closes the session, noting the competition starts in about three and a half hours and that in about six hours there will be another talk on quantum computing with Tim from NYU Shanghai. He invites participants to tune in to see what the computer that might someday help solve these problems could look like. He thanks Professor Nun Lora again, and the session ends with acknowledgments from Speaker 1.

The discussion focuses on decentralization and fears that open-source AI could be heavily censored or banned in the future, depriving people of local compute and forcing reliance on cloud systems that could be controlled. One major concern raised is “lawfare” against open-source repositories such as Z Library and Anna’s Archive. The described pattern is that large tech companies first gain access to valuable data, use it to train AI systems, and then governments intervene with legal actions that restrict access—framing the restriction as unfair—ultimately limiting what academics and individuals can use to train their own models. The result is portrayed as a situation where only large AI providers remain viable, while local inference becomes less competitive. The transcript contrasts this with China’s approach, stating China has “decided not to play this game at all” by allowing data sources to proliferate and not burning its own libraries of Alexandria. It claims that about half to two thirds of available open-source information is in Chinese, and that this could reach ninety percent. The claim is that this makes it easier to access open-source models and run them locally, including Chinese models such as Qwen and DeepSeek, which can be loaded from Hugging Face and run on a powerful machine. It emphasizes that running these models locally “won’t be able to” work on a normal gamer rig and requires specialized hardware purchased directly from Nvidia, with an example of starting around ninety-six gigabytes of RAM. The goal stated is local inference once models are available and can be run on local systems. A further concern described is a shift in political messaging: rather than stopping AI data centers, figures like Elizabeth Warren are said to be pushing for taxing people who use artificial intelligence. The transcript argues that this could become a mechanism to increase taxes while leaving people unemployed, with ongoing financial burdens. It claims that using centralized AI services such as Anthropic’s Claude, Google Gemini, and OpenAI’s Codex would mean paying the tax to “essentially only three main cartels.” The transcript concludes by describing a future enforcement model likened to marijuana interdiction, where “commissars” would ask about what is running on data servers and what inference is being conducted, and then impose taxes to regulate and charge for “cognitive labor” produced by AI models.

The discussion centers on the ongoing battle between Google and Nvidia in AI hardware, with Google focusing on TPUs and Nvidia offering a full GPU stack. Blackwell, Nvidia’s next-generation chip, faced a delayed first iteration (Blackwell 200) and was followed by a difficult, complex product transition from Hopper to Blackwell. The transition required moving from air cooling to liquid cooling, increasing rack weight from about 1,000 pounds to 3,000 pounds, and boosting power from roughly 30 kilowatts to about 130 kilowatts. The speaker likens the change to a homeowner needing to overhaul power infrastructure, cooling, and the physical environment to support a new, denser, heat-intensive system. As a result, many Blackwell SKUs were canceled, and true deployment only began in the last three or four months, with scale-out starting recently. Google is viewed as having a temporary pre-training advantage and, notably, being the lowest-cost producer of tokens. The speaker argues that, in AI, being the low-cost producer has become a meaningful factor, a rarity in tech markets. This dynamic enables Google to “suck the economic oxygen out of the AI ecosystem,” making life harder for competitors and potentially altering strategic calculations across the industry. Two key upcoming shifts are highlighted. First, the first models trained on Blackwell are expected in early 2026, with the first Blackwell model anticipated to come from XAI. The rationale is that even with Blackwells available, it takes six to nine months to reach Hopper-level performance due to Hopper’s tuning, software, and architectural familiarity. Since Hopper outperformed its predecessor after six to twelve months, Nvidia aims to deploy GPUs rapidly in coherent data-center clusters to work out bugs fast, enabling Blackwell scaling. XAI is positioned to accelerate this process by building data centers quickly and helping debug for others, thereby likely producing the first Blackwell model. Second, the GB200’s difficulties gave way to the GB300, which is drop-in compatible with GB200 racks. The GB300 will be deployed in data centers capable of handling the new heat and power requirements, replacing not the GB200s but fitting into existing, scalable racks. Companies using GB300s may become the low-cost token producers, especially if they’re vertically integrated; those paying others to produce tokens would be disadvantaged. These hardware developments have broad strategic implications for Google: if it maintains a decisive cost advantage and potentially operates AI at negative margins (e.g., -30%), it could continue to extract economic oxygen from the market and solidify a dominant position, affecting funding dynamics for competitors. The shift from training to inference with Blackwell deployments and the arrival of Rubin are anticipated to widen the gap versus TPUs and other ASICs, altering the economics and competitive landscape of AI at scale.

AMD announced the Ryzen AI Halo, a new reference platform for local AI deployment. It is described as the smallest AI development system in the world and is capable of running models with up to 200 billion parameters locally, without being connected to anything. The platform is powered by AMD’s highest-end Ryzen AI Max processor with 128 gigabytes of high-speed unified memory shared by CPU, GPU, and NPU, which AMD says accelerates system performance and enables running large AI models on a compact desktop PC. Halo is said to support multiple operating systems natively, ship with the latest “Rockham” software stack, and be preloaded with leading open source developer tools. The platform is also described as running hundreds of models out of the box, giving developers the tools needed to build, test, and deploy local agents and AI applications directly on the PC. AMD stated Halo will launch in the second quarter of this year. Greg (Speaker 0) discussed the growing compute needs required by AI, saying that building what people want to do would require far more compute than is currently available and that he would like a GPU running in the background for every person, which he described as implying billions of GPUs without a plan to build that kind of scale. He then focused on benefits in people’s lives using ChatGPT, including a healthcare story in which a coworker’s husband experienced leg pain, was initially told it was a pulled muscle, entered symptoms into ChatGPT, which advised returning to the ER and identified deep vein thrombosis and additional blood clots in the lungs. Another story described an executive, Fiji’s CEO ChachiBT, who was in the hospital for a kidney stone and infection; she asked ChatGPT about the safety of an antibiotic, and ChatGPT said it was not safe based on medical history from two years earlier, which she then showed to a doctor with limited time to review her history. Greg’s points were supported by AMD’s Speaker 1, who said AI will be everywhere over the next few years and that AI is “for everyone,” enabling people to be smarter, more capable, and more productive. Speaker 1 cited growth from about one million to more than one billion active users for AI since the launch of ChatGPT and projected over 5 billion active users as AI becomes indispensable. He attributed AI’s foundation to compute and stated demand for global compute infrastructure has grown from about one zetta flop in 2022 to more than 100 zetta flops in 2025, while adding that there is still not nearly enough compute for everything that could be done as models become more capable and as agents expand.

Speaker 0 discusses competing narratives about AI model companies, noting that some see them owning everything while others believe open source, China, or a combination of both will dominate. He highlights Kimi, which released a competitive model to the latest Claude at roughly 95% capability for a fraction of the price, illustrating the open-source/china-driven competition. He observes a notable rotation in the market: Nvidia’s sustained success over the past five years has made chips the center of action, and the stock market shows a shift from software to hardware. He asks whether chips will capture all the value and whether software will become open source, suggesting the possibility that even if chips accrue value, they might become commoditized like past tech cycles. He cautions that historically, whenever people proclaimed chips to be where the value is, they often commoditize. This leads to bigger questions about the app layer: will there be specialized apps that harness AI in areas such as medicine, where apps could be tailored and customized, or in legal and various business domains? Or will the models themselves perform all these functions without specialized applications? The speaker emphasizes the novelty of the current moment: AI is a long-standing topic (an 80-year thread), but the mode of operation now—where this set of questions is being resolved—is only partway through. He suggests we are probably in a three-year stage within a likely thirty-year shift and concedes that we do not yet know how these dynamics will unfold.

Mike Adams, executive director of the Consumer Wellness Center and founder of decentralized.tv and brightlearn.ai, recounts a costly warranty dispute over an NVIDIA RTX Pro 6,000 Blackwell Workstation Edition GPU purchased for about $9,000. He explains that the card, branded by PNY, has a faulty power bus that causes it to freeze and reboot across multiple workstations and operating systems (Ubuntu, Windows 11, various Linux distros). Adams notes he owns several of these cards and that all others in the same model perform correctly, isolating the issue to this specific unit. He describes his hardware-heavy workflow: around 48 workstations operating as part of a nonprofit data pipeline processing, including tasks like cleaning books for reference text for his book engine and search engines. He emphasizes he does not offer inference services publicly with these cards, but uses them in-house for large-scale model inference, including text, image, and video models. Adams details the warranty process, starting with contacting NVIDIA for a replacement under the three-year warranty. The sequence reveals repeated handoffs and escalating requirements. NVIDIA’s initial response required proof of purchase, photos of the card (all four sides and serial number), a photo of the workstation, and then a photo of a handwritten case number next to the serial number. He then provided a full system dump using a Windows utility, which was sent to NVIDIA. The process supposedly moved to a replacement team, which again requested proof of purchase, more photos, and additional utilities to run. Despite compliance, he was told to contact the reseller rather than NVIDIA. Assurant Technologies, the reseller based near Dallas, was then involved. Adams reports that Assurant required him to download and run a utility named Extern SWAC, allegedly from Google Drive, and to rename it with a .exe extension and run it as administrator. He cites BraveSearch identifying Extern SWAC as malicious, a security tool that purportedly performs VM detection, hides debugging tools, and modifies registry keys. He refused to download or run this file, asserting it could compromise his system. He offered to provide telemetry analysis scripts (ClaudeCode) to recreate the failure instead. Sheng Shu of Assurant allegedly forwarded the case to PNY. Adams then engaged with PNY’s technical support supervisor, Bruce P, who requested additional proof of purchase and the execution of further tools. Adams had already supplied multiple proofs of purchase, serial numbers, and extensive telemetry reports, including two test reports and a crash analysis indicating hardware defects. He presented a detailed telemetry package showing: 216 driver errors, five BSODs, zero ECC errors, and VBIOS corruption, with a conclusion that the root cause was a hardware defect in the GPU’s power delivery VRM subsystem. The ClaudeCode analysis described an abrupt termination with a hardware-level failure, not software degradation, and recommended RMA. PNY allegedly rejected the case, insisting that Adams run another utility and accept more steps, even after extensive evidence. Adams states that he refused to run what he views as malware and that PNY would not honor the three-year warranty, instead passing responsibility through NVIDIA, Assurant, and then back to PNY. The outcome, according to Adams, is a warranty scam: he claims a defective card has not been replaced, and the three-year warranty is not honored. He asserts that this behavior is fraudulent and warns consumers not to buy NVIDIA or PNY products, stating that they will not honor warranties and may even compel customers to install malware as a condition of service. He says he has filed complaints with attorneys general and consumer boards and suggests alternatives like Intel, AMD, and Apple for GPUs and unified memory solutions. He ends by reaffirming that this experience with NVIDIA and PNY is a cautionary tale for consumers.

- Gavin Baker is deeply engaged with markets beyond his quantitative investing background, with a passion for technology investment and wide-ranging views on NVIDIA, Google and its TPUs, the AI landscape, and the evolving business models around AI companies. He even entertains ideas like data centers in space, arguing from first principles that they are superior to Earthbound data centers. - The host and Baker discuss how to process rapid AI updates (e.g., Gemini 3). Baker emphasizes using new AI tools personally, paying for higher-tier access to get mature capabilities, and following leading labs (OpenAI, Gemini, Anthropic, xAI) and influential researchers (e.g., Andre Karpathy). He notes that AI progress is heavily influenced by public posts and discourse on X (formerly Twitter), and highlights the importance of embedded signal from the lab ecosystem and industry insiders. - On Gemini 3 and scaling laws, Baker argues that Gemini 3 affirmed that scaling laws for pre-training are intact, an important empirical confirmation. He compares the public’s overinterpretation of free-tier capabilities to that of a ten-year-old, stressing the need for paying for higher-tier capabilities to gauge real performance. He explains that progress in AI since late 2024 hinges on two new scaling laws: post-training reinforcement learning with verified rewards (RLVR) and test-time compute. He emphasizes that these laws enable better base models and that Google’s TPU strategy and Nvidia’s GPU strategy each shape the competitive dynamics. - Baker details the hardware race between Google (TPUs) and Nvidia (GPUs), including the transition from Hopper to Blackwell as a massive product shift requiring new cooling, power, and architecture. He credits “reasoning” (and reasoning-based models) with bridging an eighteen-month gap in AI progress, enabling continued improvement without the immediate need for Blackwell-scale infrastructure. He explains that Blackwell deployment has been slower but is now ramping in significant fashion, and that RBMs (Blackwell clusters) are likely to dominate training eventually, with current GB-300 and MI (Mixtures) chips enabling future efficiency gains. Rubin, as the next milestone, is anticipated to widen the gap versus TPUs and other ASICs. - Google’s strategic move to be a low-cost token producer is highlighted as a way to “suck the economic oxygen” out of the AI ecosystem, pressuring competitors. Baker predicts first Blackwell-trained models from XAI in early 2026, and posits that Blackwell will not immediately outperform Hopper but will be a superior chip once fully ramped. He discusses TPU v8/v9 as potentially high-performance but notes Google’s conservatism in design decisions and their reliance on Broadcom for backend manufacturing. He foresees a shift toward in-house semiconductor development eventually as the cost and margins of external ASICs become less attractive. - The potential shift to in-house semiconductor production is tied to economics: if token production scales and external margins (Broadcom) are too high, Google could renegotiate or internalize more of the stack. This would affect margins and the competitive landscape, including whether Google remains the low-cost producer. - In discussing broader AI deployment economics, Baker notes the importance of inference ROI, with concerns about an initial “ROIC air gap” during heavy training phases. He cites CH Robinson as an example of AI-driven uplift in a Fortune 500 company, where AI enabled 100% pricing/availability quoting in seconds, boosting earnings. This example supports the view that AI-driven productivity improvements can boost profitability even as capital expenditure remains high. - Baker discusses the outlook for frontier models and the likely near-term impact on industries, including media, robotics, customer support, and sales. He suggests that the most valuable AI systems will rapidly become useful and context-aware, capable of handling long context windows (for example, by remembering extensive user preferences) and performing complex tasks like travel planning or hotel reservations. - On the economics of AI-driven product development, Baker argues that AI-native SaaS companies must accept lower gross margins to achieve ROI through much higher efficiency and automation. He contrasts this with traditional SaaS margins, noting that AI enables substantial gross profit dollars through reduced human labor, while demanding reinvestment in compute. He urges traditional software companies to embrace AI-enabled agents and to expose AI-driven revenue streams, even if margins are compressed. - Baker reflects on the broader tech ecosystem, including private equity’s potential to apply AI systematically, and the role of private markets in scaling semiconductor ventures. He emphasizes that AI requires an ecosystem of public and private players across chips, memory, backplanes, lasers, and more, and that China’s open-source efforts may be insufficient to close the gap created by Blackwell’s advancement, given the looming lead of U.S. frontier labs. - The conversation also touches on space-based data centers as a transformative, albeit speculative, frontier: advantages include perpetual sun exposure for power, reduced cooling needs, and ultra-fast laser-linked interconnects in space. The main frictions are launch costs and the need for new infrastructure (Starships, global collaborations), but the potential synergy with AI hardware ecosystems (Tesla, SpaceX, XAI, Optimus) is noted as strategically significant. - In closing, Baker emphasizes that investing in AI is the search for truth, with edge coming from uncovering hidden truths and leveraging history and current events to form differential opinions. He attributes his own lifelong motivation to competitive drive, a love of history and current events, and a relentless pursuit of understanding the world’s technology and markets.

Jensen Huang (NVIDIA) discusses how the amount of compute—and the energy required for that compute—is likely to increase dramatically, moving from “a hundred times” to “a thousand times” compared with current levels. He frames future computing as two simultaneous shifts: it will be intelligent and contextually aware with generative outputs, and it will be continuous rather than based on prerecorded retrieval that is initiated only when prompted. The discussion contrasts concerns about today’s AI being “backward looking” and copying previous work, potentially leading to feedback loops where people rely on AI and become stagnant without new regenerative creativity. Jensen Huang’s described future addresses this by arguing that software will not remain static code stored on a hard drive; instead, people will ask AI to write software in real time as needed (for example, generating a Photoshop clone to edit an image or generating an original movie tailored to a preference). Creating such continuous generative experiences is said to require a tremendous amount of energy—“a thousand times more” than today’s levels. Speakers note that existing energy sources cannot easily support this scale. The conversation states that it cannot be done on hydrocarbons, not even on nuclear due to long build-out time, and not on solar because current energy sources are insufficient. It also emphasizes efficiency: having the ability to use vastly more energy does not mean it should be used, and continuous regeneration is not always the more efficient approach. Speaker 0 then argues for limiting market cap and having these groups invest themselves without government backing or government liability protection, suggesting a free-market approach rather than government-directed competition framed as an arms race. Speaker 2 responds that pursuit of “superintelligence” requires centralized power and therefore cannot be decentralized. The conversation claims this centralized effort is being directed toward a quest for superintelligence connected to world domination and competition, particularly framed as an attempt to “beat China,” and concludes that once superintelligence is achieved, humanity’s fate would be in question.

The speaker reframes computers as AI factories, which produce tokens, numbers. These AI factories should be used for three fundamental things, with the first being to train the next frontier model so you can build the best AI and get to market first. The goal is to train it as fast as possible. Regarding performance, Rubin is described as a 4x leap compared to Blackwell, meaning the fourfold improvement could be achieved in one month instead of four months.

The speaker argues that China’s export restrictions on indium compounds will likely disrupt the AI data center infrastructure build-out. They describe themselves as an AI developer and “elemental scientist,” running a mass spec laboratory for elemental analysis, and emphasize indium’s rarity and lack of natural abundance. They connect indium to periodic table groupings: indium is in the same periodic table column as boron, aluminum, and gallium (group 13), and like other group 13 elements it has three outer electrons. They state that combining a group 13 element with a group 15 element such as phosphorus produces compounds with characteristics “like silicon” but with better light transmission. They assert that for systems requiring optical transparency and conductivity—solar panels, optoelectronics, optical telecommunications, touch screens, solar cells, and electrodes embedded in displays—indium enables transparent conductors such as indium tin oxide. The speaker links indium to high-speed data center networking, claiming copper cannot provide the required throughput for massive GPU clusters (they mention setups like 100,000 GPUs). They say extremely fast GPU-to-GPU interconnections require optoelectronics and optical transmission rather than copper wiring, noting that they personally use copper at 10G but that it is “getting really slow,” while large AI builders (SpaceX, OpenAI, Meta, Google) rely heavily on optical infrastructure. They claim data centers thus “depend severely on indium.” They then describe escalation in export controls: China is restricting indium exports (and “scrutinizing” exports of straight indium). They say that even in 2025 China added indium phosphide to an export control list. They explain that indium phosphide is indium and phosphorus configured together. They state that indium phosphide wafer prices increased by about 250% in a little over a year and a half, reaching about $5,000 per 6-inch wafer, and they portray this as the “template” for optoelectronics fabrication. The speaker further claims that China is asking extra questions of buyers of just indium, including European and U.S. purchasers providing end user information and destination country details. They connect this to prior U.S. pressure on ASML to block high-end UV lithography exports to China and say China is countering by blocking gallium exports and indium/indium phosphide exports. They argue this will “dramatically hamper” U.S. AI data center build-out, stating that silicon does not work at required wavelengths while indium phosphide works for lasers, photodetectors, modulators, and optical telecom equipment for terabits-per-second bandwidth. They claim “there is no substitute in photonics” for indium phosphide and state that without indium there is no high-speed optical networking in data centers. They present supply chain choke points: they say China controls about 70% of the global indium market and also point to AXT Sumitomo as handling about 80% of substrate manufacturing, while non-Chinese buyers depend on China-controlled input. They reference a Mining.com story stating China’s control over indium phosphide exports threatens AI data center rollout and quotes Semi Analysis analyst Conrad Wong on indium phosphide as a supply chain bottleneck gating AI data center build-outs. They mention NVIDIA’s $2 billion investment into U.S. photonic product makers Coherent and Lumentum and Marvell acquiring Celestial AI, claiming these moves reflect an industry need for photonics dependent on indium. The speaker expands to related shortages and production constraints, mentioning gallium and tungsten hexafluoride (WF6) as bottlenecks for microchips and optoelectronics. They explain indium comes as a byproduct from zinc mining rather than from dedicated indium mines, stating there are “no dedicated indium mines” and that indium is extracted from zinc ores using solvent extraction and electro-refining. They claim China mines/refines around 70% of indium supplied globally, followed by South Korea, Japan, Canada, and others, and state none is the United States. They assert that while indium recycling exists (especially reclaiming indium tin oxide from displays in Japan), there are “almost no spare reserves,” and they say there is no U.S. mining or large-scale U.S. reclamation sufficient for AI data centers. They conclude that if China “flick[s] a switch” to block exports, the U.S. AI industry could be stopped quickly due to dependence on Chinese supply, and they argue that without indium there is no quick substitute. They add element trivia, stating indium is named from “indigo” due to its bright indigo blue spectral line and the Latin indicum, and they mention other elements as named after places or scientists. They end by urging caution toward AI company hype, warning that AI data center expansion could hit “a brick wall called no indium,” tied to ongoing export restrictions and supply bottlenecks.