Gavin Baker - AI, Semiconductors, and the Robotic Frontier - [Invest Like the Best, EP.385]
What this covers
Patrick O'Shaughnessy talks with Gavin Baker, CIO of Atreides Management, who has followed Nvidia since the turn of the millennium. The conversation starts with how AI training actually works: coherent clusters in which GPUs share memory, the largest recently being 32,000 H100s, of which only about 15,000 to 16,000 could run at once. Baker proposes MFU as the best single metric for judging a lab, with typical figures of 35 to 40 percent. He illustrates software maturity by comparing Nvidia's roughly 83 percent with AMD's MI300 and MI250.
The middle stretch is about the spending race. Baker holds that scaling laws remain intact, that the apparent slowdown since GPT-4 reflects the wait for Blackwell, and that clusters of around 300,000 GPUs will be the real test. He argues that ROIC has risen at the big spenders and that Meta's reaccelerated ad growth shows returns already. He frames the rivalry as a prisoner's dilemma driven by a race to a "Digital God," from which Apple is exempt because it monetizes through iOS distribution. Contested claims include GPT-7 or GPT-8 costing around $500 billion and Larry Page reportedly being willing to go bankrupt rather than lose.
The conversation then widens. Topics include Tesla's step-function FSD gains, a possible small on-car LLM within 6 to 18 months, and the risk that synthetic video data fails. Baker also predicts humanoid robots will beat specialized ones, and that solar and batteries will compound until fossil fuels disappear. On investing, he argues Reg FD erased hedge funds' access edge and quants removed value's behavioral alpha. He also asks whether fundamental investors armed with LLM-built databases might take share back from quants.
Baker argues AI economics are defined by very high marginal cost, so the winners will be whoever compounds the most compute efficiency (MFU decomposed into software, system, checkpointing and power efficiency) and combines it with unique data and distribution, with inference increasingly shifting to phones.
- Scaling laws mean quality improves only with much more spending, unlike software's zero marginal cost
- MFU decomposes into MAMMF x SFU x checkpointing x PUE, and small percentage gaps compound into large cost and speed advantages
- Models converge in capability, so durable advantage comes from unique real-time data, distribution, and local inference on devices
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Model FLOPS utilization (MFU), the share of theoretical compute actually applied to training, runs around 35 to 40 percent, with published GPT-3, LaMDA and Megatron figures in the high 20s to high 30s, and Baker argues it is the best single metric for evaluating a lab.
“this has been measured today in something called MFU, model FLOPS utilization, and that generally runs around 35% to 40%.”
Training requires a coherent cluster in which GPUs share a memory space so they effectively know what each other is computing; the largest coherent cluster until recently was 32,000 H100s, of which only 15,000-16,000 could run at once due to efficiency problems.
“The biggest coherent cluster in the world until very recently was 32,000.”
Humanoid robots will beat specialized robots because a general-purpose form factor can do any human task in a world built for humans and will reach manufacturing scale that gives it a cost advantage, the way GPT generalized AI, and the physical world will be reshaped around them.
“humanoid robots are the future, not the specialized robots.”
Reg FD and transcript availability ended most information advantages for hedge funds, which have no access edge over large long-only firms like Fidelity or T. Rowe Price, whose access is orders of magnitude greater.
“from my perspective, Reg FD means there's almost no reason to talk to companies”
Value strategies once earned alpha from human emotion, such as embarrassment and career risk aversion, but algorithmic quant investing removed that edge, which is why there is little dispersion within the value factor today.
“the value strategy is the alpha from them, was based on human emotion, and people being embarrassed to buy stuff”
Leading labs will keep spending heavily until there is irrefutable evidence that scaling laws are slowing, and the lack of a new Nvidia GPU generation is why progress appears to have slowed since GPT-4, since the next generation is needed for the next step-function jump in capability.
“the only way you get irrefutable evidence why has progress slowed down since GPT-4 came out is there hasn't been a new generation of Nvidia GPUs.”
Meta's stock fell about 80 percent partly from metaverse overspending and partly from Apple's IDFA changes that removed targeting, but its revenue growth has reaccelerated because AI-driven targeting and creative optimization restored ad performance, so AI ROI is already visible.
“Meta went down like 80%, partially because they're overspending on the Metaverse, but partially because Apple took away their ability to target through IDFA.”
Fundamental investors who combine deep domain knowledge with LLM-powered vector databases of their own data may take alpha share from quantitative investors over five to ten years, a hypothesis Baker admits could fail.
“the only alpha left in the market for fundamental investors at the edge of probability”
Tesla's miles-between-disengagements improved in step functions with FSD v12 and 12.5 on AI4 hardware, and Tesla alone has a visual driving dataset based on miles driven that may be 100x or more larger than Waymo's, making it the first robot likely to affect the world at scale.
“I think the first robot that's really going to impact the world is every Tesla car”
Putting a speed-optimized small LLM with a world model onto each Tesla could provide enough reasoning to handle novel driving situations, unlocking another step-function jump in FSD capability within about 6 to 18 months, though synthetic video data's effectiveness is unproven.
“if you put a really speed-optimized small LLM in -- locally on each Tesla, there might be just enough reasoning capability to unlock another step function”
Commodity-seeming frontier models will not stay commodities if scaling continues, because training GPT-7 or GPT-8 may cost around 500 billion dollars, making scale the most powerful barrier to entry.
“GPT-7 or 8 literally costs $500 billion to train, I don't think they're going to stay commodities”
ROIC has risen at the big AI spenders since they ramped CapEx because NOPAT grew faster than CapEx as companies trade human labor for GPU hours, so the AI ROI skeptic case is contradicted by the financials.
“ROIC has gone up for all of these companies since they ramped CapEx.”
Next-generation Blackwell clusters of around 300,000 GPUs should yield a GPT-5 to 6 class model; the only way to disprove scaling laws is a 10x compute increase from new GPUs and networking that fails to produce a major quality improvement.
“if you get that 10x the order of magnitude improvement of compute, and you don't see a massive improvement in model quality, then scaling laws will stop”
Photovoltaic efficiency compounds just under 10 percent a year and battery efficiency compounds about 200 basis points less, so economics will make sunlight-based power, including stored sunlight, cheaper than other sources and fossil fuels will go away.
“photovoltaic cell efficiency is compounding just under 10% a year, and battery efficiency is compounding 200 bps less than that.”
Synthetic data appears to improve model training even though no one understands why, and whether it will keep working is unknown, which makes it a key uncertainty for future model scaling.
“no one understands why, but it looks like it works. Now again, will it continue working? I don't know. Nobody knows.”
Larry Page has reportedly said internally at Google that he is willing to go bankrupt rather than lose the AI race.
“Larry Page has evidently said internally at Google many times, "I am willing to go bankrupt rather than lose this race."”
The biggest risk to Tesla's autonomy thesis is that synthetic video data does not work as synthetic written data does, and regulators will demand a very high safety bar for AI driving.
“We don't know if synthetic video data works. Nobody knows.”
MFU decomposes into MAMMF (software efficiency) times SFU (system FLOPS efficiency) times the share of time not spent checkpointing, and then times PUE (power utilization efficiency), so these factors compound multiplicatively into the final efficiency of a training system.
“you need to multiply MAMMF times SFU, system FLOPS efficiency, and then you need to multiply that by the percentage of time that is spent at a checkpoint per training run.”
Big Tech leaders such as Zuckerberg, Nadella and Pichai are not focused on AI ROI because the founders who control these companies believe they are in a race to create a Digital God, and losing that race is an existential threat, so ROI is secondary to winning.
“the people who actually control these companies, the founders, there's either super-voting stock or significant influence in the case of Microsoft, believe they're in a race to create a Digital God.”
MAMMF (maximum achievable matrix multiplication FLOPS) measures software efficiency relative to theoretical chip peak: Nvidia GPUs reach about 83 percent, AMD's MI300 rose from about 25 to 30 percent to about 60 percent, and the MI250 reached only about 10 percent, showing that software maturity determines whether hardware is usable.
“Nvidia GPUs run it, per his testing, 83%.”
The true metric is exaFLOPS per second per dollar of CapEx per watt of electricity, and labs that optimize it could reach more than a 100 percent advantage, separating a GPT-7 or 8 that costs around 1 trillion dollars from a next class that costs 200 to 400 billion dollars.
“some of these companies are going to have greater than a 100% advantage at exaFLOPS per dollar of CapEx per watt consumed.”
The value of AI models comes from distribution and unique data rather than the model itself, which is why Meta open-sources Llama and xAI open-sources Grok-1 while Google can monetize its proprietary data such as YouTube and Maps.
“the reason is the value may not come from the model.”
Unless someone builds a compounded efficiency advantage, frontier models will converge to roughly the same intelligence, so the decisive differentiator becomes who has the most differentiated real-time data combined with Internet-scale distribution.
“I think they're probably all going to converge to roughly the same intelligence.”
Compute alternates between centralization and decentralization; having reached the end of a long cloud-centralized period, inference is moving to the edge, with phones becoming the main inference platform.
“compute tends to cycle in between centralization and then decentralization.”
Rival labs face a classic prisoner's dilemma in which a mutual agreement to slow Blackwell cluster buildouts would be a Nash equilibrium, but the race to digital god makes that unreachable, while Apple avoids the trap because it monetizes AI through its iOS distribution toll booth.
“All these other companies are trapped in this prisoner's dilemma.”
Inference on the device is effectively free while cloud inference costs money because it burns GPU hours, so the rule of thumb becomes local when you can, cloud when you must, which will push workloads onto phones.
“it will be local when you can, cloud when you must. So if you could inference locally on your phone, you will always do that because it's free”
Investors should target the most constrained or inefficient link in the AI efficiency chain, such as low SFU, checkpointing reliability, or PUE, because improving a binding constraint produces outsized returns.
“in tech, whenever you find a constraint, if you could invest in something that alleviates that constraint, you usually do well”
Standard SaaS benchmarks of about 1 million dollars ARR in year one, 5 million in year two and 10 million in year three became quantitative shortcuts that inflated multiples and overfunded verticals, leaving many 2021 SaaS investments impaired.
“which everybody view is one reason SaaS multiples got so inflated because it became almost quantitative.”
Unlike software with zero marginal cost, AI has very high marginal costs because scaling laws mean quality improvements require spending substantially more, so the economics resemble infrastructure more than software.
“AI has extremely high marginal costs because scaling laws literally mean that the only way you get an improvement in quality is by spending a lot more, full stop.”
Modern corporate hierarchies reward charisma and political skill, producing CEOs whose egos grow faster than their competence, which is why exceptional leaders like Elon Musk and Jensen Huang differ from most CEOs.
“the modern American corporation in the same way like almost any large organization with its hierarchies is set up to reward people who are charismatic and political.”
Investing has undergone successive meta-shifts, including ROIC over ROE, Reg FD, credit card data, and expert transcripts, and LLMs are likely the next and biggest.
“Credit card data was a big meta shift. Expert transcripts have been a big meta shift. I think LLMs are going to be the biggest meta shift.”
The CUDA advantage can be quantified: Nvidia reaches about 83 percent of peak per GPU, AMD's MI250 reached only about 10 percent so nobody used it, and the best labs reportedly run near 95 percent in their own optimized stacks.
“They can run at 83%. AMD is running at 25%”
Power usage effectiveness is the final multiplier, and labs can trade off networking spend against power efficiency because optimizing system FLOPS may raise PUE, so the real metric is exaFLOPS per second per dollar of CapEx per watt.
“we want exaFLOPS per second per dollar of CapEx per watt of electricity”
Future phones will be sold on DRAM rather than flash storage because DRAM determines the parameter count of the local model that can run, which determines the quality of on-device intelligence that can access private data.
“What you will most care about is the amount of DRAM you have on your phone because that will determine the perimeter count of the model you could run local.”
Companies whose CEOs want bad news delivered immediately and bypass hierarchy tend to perform better, while bankrupt companies often share a CEO who did not like hearing bad news.
“the number one thing all bankrupt companies had in common was the CEO, who didn't like to hear bad news.”
People with better local or cloud AI will gain large advantages, potentially 15 to 20 IQ points, creating an inequality of intelligence access that Baker finds dystopian but thinks is likely.
“I'm going to be really advantaged as a human. And that seems dystopian to me, but it's also hard for me to not see that happening.”
ChatGPT in 2022 is to AI what Netscape Navigator in 1994 was to the internet, and less than 1 percent of current internet market cap was founded in the two years after Navigator, so most value-creating AI companies have not yet been founded.
“only less than 1% of current global internet market cap was founded in the two years after Netscape Navigator came out.”
A higher MFU lets a lab choose among faster time to market (45 percent faster at 50 versus 40 percent MFU), higher quality, or lower cost; a 25 percent higher-quality model combined with quantization can cut inference cost by almost 50 percent.
“It's almost like if you have a 25% higher-quality model because you have a higher MFU and you build in quantization, then you can actually get almost a 50% reduction in inference cost”
Over the last five years GPUs have gotten roughly 50x faster while the rest of the data center (networking, storage, memory) got only 4x to 5x faster, so GPUs sit idle waiting for data and MFU stays low.
“GPUs have gotten 50x faster and the rest of the data centers only got 4x to 5x faster. And that is why MFU is so low”
Compounding moderate efficiency gaps produces large output differences: a firm at 90 percent MAMMF, 50 percent SFU and checkpointing one-third of the time reaches about 30 percent MFU, while a firm at 100 percent MAMMF, 60 percent SFU and 10 percent checkpointing reaches about 54 percent.
“if you have to checkpoint only 10% of the time, you're at 54%, your competitor is down at 30%.”
Thin AI wrapper companies must build defensibility through a strong sales motion, easy integration points, RAG on messy customer data, and avoiding extensive fine-tuning that locks them to a model generation, and they should build compound AI systems that route queries to the cheapest adequate model.
“How are you creating a compound AI system such that you're serving each query at the lowest cost possible”
Jensen Huang and Elon Musk have no fixed schedules or standing meetings; they go wherever the most important problem is, pull the best people onto it, and do not require hierarchy, which is a key reason their companies attract exceptional talent.
“I just find out what is the most important problem at the company, and I go, and I sit my desk out in that area”
Baker concentrates investments on companies that relieve the current binding constraint in AI infrastructure, which he believes is SFU, checkpointing reliability, and PUE.
“in tech, whenever you find a constraint, if you could invest in something that alleviates that constraint, you usually do well, and right now, I profoundly believe that constraint is at SFU, checkpointing, which goes to reliability and then PUE.”
xAI built a 100,000-GPU coherent training cluster in Memphis by redesigning the data center from first principles, achieving enough density to train on Hopper GPUs without next-generation networking from Nvidia or Broadcom.
“I think he rearchitected the xAI data center from first principles in Memphis. It's very different than other data centers.”
Only about three places in the United States can reliably supply one gigawatt of power to a single data center, essentially nuclear sites; average electricity costs about 8 cents per kilowatt-hour, but a gigawatt is priced around 10 times that, which makes PUE economically significant.
“There's only basically three places in the United States today where you can get 1 gigawatt of power to a single data center that's reliable enough, i.e., that's nuclear.”
Apple will monetize AI by using a router to direct queries to the best cloud model per dollar and charging AI companies increasing tolls to be included, and will offer tiered subscriptions for cloud intelligence, similar to how Google collected tolls from search.
“Apple will charge companies to be in their router, and then, oh, if you want to be routed too, you're going to have to pay a bigger and bigger toll.”
AI-first application companies are growing from zero to 30 million dollars in nine months and, despite high marginal costs, can be more cash-flow efficient than software companies, and they are attacking software and labor budgets in nearly every vertical.
“there's all these companies that are growing like $0 to $30 million in 9 months.”
LLMs will be the biggest meta shift in investing, but they may hit purely quantitative investors hardest in the next five to 10 years, creating a window where fundamental investors can recapture alpha from quants.
“I think LLMs are going to be the biggest meta shift.”
As LLMs democratize domain knowledge, the scarce skill in early-stage venture becomes judgment quotient (JQ), particularly assessing team quality, with knowledge quotient (KQ) and emotional quotient (EQ) as secondary.
“it will really for a while place the emphasis on JQ, judgment quotient.”
Frontier models will likely converge to similar intelligence unless a lab builds a compounding advantage in MFU, SFU, checkpointing, and PUE, and firms like Google and Meta will try to close gaps by spending more money.
“I think they're probably all going to converge to roughly the same intelligence.”
Large amounts of undifferentiated white-collar work can be replaced now by LLM tools that return analyst-quality output almost instantly and for near-zero cost, as Patrick's small team and Baker's firm's intern experience show.
“how often I use this thing in a way that I normally would have used an analyst, it's multiple times a day, and it's returning work as good.”
Firms like Renaissance and Two Sigma can pay tens of millions per year to top AI researchers and hold unique data, so LLMs may be hardest for purely quantitative investors to exploit in the near term.
“They are the only people who can compete with these tech companies who can throw $20 million to $30 million or $40 million or $50 million at a leading AI researcher per year.”
Application software sells efficiency to humans, so when AI replaces rather than augments seats, per-seat SaaS revenue declines and incumbents face a problem that they are already aware of.
“if it's replacing humans, if you're selling application software at a per seat model and those seats start to go down, you have a problem.”
Late-stage venture and growth equity are converging on private equity dynamics, where deals are auctions with no sourcing or pricing advantage, so differentiation shifts to real operational value add, as Valor provides.
“Everything is an auction. Literally, every deal that these big firms do is an auction.”
Investing in next-generation networking, storage, and memory is sensible because reaching a million-GPU cluster with Rubin-generation chips requires breakthroughs at every step, otherwise MFU could fall to 3 to 5 percent.
“Otherwise, it's going to be wasted and MFU is going to be 3% to 5%”
Even though a single Nvidia GPU runs at 83 percent efficiency, cluster-level MFU falls to about 35 to 40 percent because system FLOPS efficiency, covering networking, storage, and memory, captures the losses that happen as data moves to the GPU.
“the reason is something I would call SFU, system FLOPS efficiency”
Training clusters must be fully coherent, so a single GPU failure loses everything since the last checkpoint; GPUs and their networking, storage and memory fail often, so labs checkpoint frequently, and more reliable networking or cooling reduces how often checkpoints are needed.
“If any one GPU fails, you lose everything for the last time you saved the model, which was called the checkpoint.”
Giant AI data centers could be sited in remote shale gas fields with on-site power plants, far from cities, as an intermediate-term answer to the shortage of nuclear power in America.
“I think we'll eventually see giant data centers in shale gas fields with power plants”
Investing at the AI application layer feels very hard today because of the Chase Coleman statistic showing most value-creating companies come years later, so Baker is deliberately cautious there while acknowledging he may miss outsized outcomes.
“I think the application layer is really, really hard.”
Infrastructure efficiency and excellence will be the single most important success factor for model companies because high marginal costs make utilization of compute the main differentiator.
“infrastructure, efficiency and excellence, I think, is going to emerge as the single most important success factor”
Because rivals are waiting for Blackwell to build 100,000-GPU clusters, xAI has arguably about a one-year advantage, which is untenable for other labs, and they are now scrambling to build their own clusters without first-principles data center design.
“xAI is going to have arguably a 1-year advantage, which is untenable to these other labs.”
About a decade ago Jensen Huang recognized the importance of researchers like Geoffrey Hinton and Yann LeCun and committed Nvidia to the view that intelligence is an engineering problem solved by gluing more GPUs together, which became the company's mission.
“he started bringing out people like Geoffrey Hinton and Yann LeCun.”
AI follows the same pattern as Vaclav Smil's energy history, in which each new prime mover is followed by long gains from efficiency, as coal turbines improved from capturing 10 to 15 percent of energy to around 98 percent.
“the early days of coal, the turbine only captured 10% or 15% of the available energy from coal itself and now we're at like 98%”
Robotics like GPT will advantage incumbents that have data, compute and capital, while most non-humanoid robotics startups are unlikely to become Google-scale companies, and manufacturers with battery, actuator and motor expertise are favored.
“I do think incumbent manufacturers who have expertise and battery design, actuators, motors with big datasets are going to be advantage.”
Crowdsourced miles-between-disengagement data shows FSD made step-function improvements when Tesla cut over to end-to-end deep learning, to version 12.5 on AI 4 hardware, and later to 12.3, with the 12.3 release containing the progress of the prior decade.
“when they cut over to WAF 3, which is effectively all deep-learning, I think, eliminated almost all human code, something dramatic changed”
A data center works like a restaurant: the GPU is the head chef, storage is the delivery truck, PCIe is the worker moving food into the fridge, DRAM is the fridge, the CPU-to-GPU link is the sous chef, and GPU memory is the stove, so a GPU without memory, storage and networking can do nothing.
“the primary unit of compute in an AI server, it is the GPU, is the head chef.”
The metaverse will arrive first through augmented reality and then through brain-computer interfaces, which may become the next compute platform after phones.
“BCI may be the next true way we interact with computers after phone”
Baker expects a GPT-4.5-class model within six to nine months from xAI's training, followed after Blackwell by 300,000-GPU clusters and then GPT-5 to 6 class models.
“the first GPT-4.5 class model sometime, I don't know, in the next six to nine months”
Inference is cheaper and simpler to optimize than training because it depends mainly on memory bandwidth and on-chip memory, and it directly shapes user experience through tokens per second, which matters for search UX.
“Inference is just a lot easier because it really, really just comes down to memory bandwidth and on-chip memory.”
Google's aggressive price cuts around 2014-2015 were matched by Amazon, which impaired Amazon's cloud revenue growth and fed fears that cloud computing would become a commodity, a fear Baker says looks ridiculous now.
“Google cut prices aggressively in something like 2014, 2015. Amazon matched”
AI lets ad platforms generate and optimize around a million creatives in real time, replacing human-driven demographic and timing guesses in advertising.
“we're going to have 1 million creatives that's going to be optimized on the fly”
Building a gaming PC is the best way to understand that every computer, from iPhone to data center, has four core elements: compute, memory, storage, and the networking that connects them.
“every computer, whether it's the iPhone, whether it's a laptop, whether it's a data center, has the same, what I would call, four fundamental elements”
Modern corporate hierarchies reward charismatic, political people, who often become CEOs with inflated egos and poor judgment, so the effective decision-makers are frequently a key lieutenant behind the CEO.
“the modern American corporation in the same way like almost any large organization with its hierarchies is set up to reward people who are charismatic and political.”
Companies with a messianic mission orientation, such as SpaceX, Nvidia, Tesla and xAI, attract exceptional employees who want to work directly on critical problems, so they can operate with little hierarchy and far fewer employees than peers.
“So I think that mission orientation, there's no need for hierarchy. You get really exceptional employees.”
It would be dystopian for the world to converge on one dominant AI model, because its values would be imbued across humanity; competing models with different value systems are needed to protect objective truth in an era when it is under attack.
“I think it's supremely important for humans that we do not end up in a world where there is just one dominant model.”
Humans of average intelligence who use AI can perform like much higher-IQ humans, so the most competitive humans will be those with the best AI agents, and advantage will compound as AI improves.
“there's a lot of humans with -- I don't know I have no idea... a lot of people are missing is that there's a lot of humans”
Internal AI tools are replacing white-collar scutwork, the hard unpleasant work Jack Welch described, and usage rises quickly from a couple of hours a day as users learn workflows, which means much undifferentiated labor will be automated.
“all of these little wrappers or mini wrappers, whatever we are going to call them, are going to replace a lot of that scutwork.”
AI assistance can lift an average human's effective capability substantially, for example from about 100 IQ to about 115 when using AI, which explains why many skeptics underrate AI's impact.
“So let's just say it's 100 IQs, and they can use AI and all of a sudden, they're like a 115.”
In semiconductors it usually takes until a chip's third generation to hit peak performance because software must mature, as with Google TPUs, which is why AMD took so long to break into the AI accelerator market.
“it almost always in semiconductors, it almost always takes you till your third-generation chip to really, really hit it.”
Reading transcripts is far higher bandwidth than listening, perhaps five to 10 times faster, so reading expert transcripts revealed insights that Baker missed in conversation, which he sees as a major meta shift.
“my input speed for reading is probably something like 5 to 10x faster than humans could talk.”
Some leading labs may run CUDA more efficiently than Nvidia itself, making their software optimization a source of proprietary advantage.
“a lot of people have said some of these labs know how to use CUDA more efficiently than Nvidia, that's a big secret sauce.”
Major tech companies were in discrete competitive swim lanes for a long time, overlapping only in cloud, which was a stable oligopoly; generative AI, as a general-purpose technology, put them all in the same swim lane and made them feel existential threats to each other.
“It's such a general technology that they're all of a sudden in the same swim lane and they all feel like it's existential.”
Apple's advantaged position comes from moving inference onto devices, which is clearly its strategy, while other hyperscalers remain caught in competitive races for cloud compute.
“So this is clearly Apple's play. It is one reason that they are in such an advantaged position.”
Re-reading company transcripts after a year off showed that management rarely says anything not already in the transcript, making transcript analysis a valuable research method.
“they never say anything that is not in a transcript.”
In an illustrative example, a company with 60 percent MAMMF and 60 percent SFU that checkpoints only 10 percent of the time reaches about 54 percent utilization, while a competitor at 45 percent SFU checkpointing a third of the time falls to about 30 percent.
“if you have to checkpoint only 10% of the time, you're at 54%, your competitor is down at 30%.”
Investors can have a direct impact on the world by lowering AI infrastructure costs through networking, storage, and utilization software, and by funding labs that compete with Google and Meta, which supports the goal of plural AI.
“as investors, by lowering and improving the cost of AI infrastructure”
Five labs now have GPT-4-class models, namely Google, OpenAI, Anthropic, xAI, and Meta, and MFU is the best single metric to evaluate which of them is succeeding among them.
“there's been five GPT-4 class models from – it's Google, OpenAI, Anthropic, xAI and Meta”
Kevin Scott said he has seen early GPT-5 checkpoints and that scaling laws are continuing, which Baker considers the best available indication that scaling is still working.
“Kevin Scott just did a podcast and he basically said, Look, I've seen some early checkpoints of GPT-5s, and scaling laws are continuing.”
Only a few companies, including xAI, Google, Microsoft, OpenAI via Microsoft, Amazon, Anthropic and Meta, combine unique data with Internet-scale distribution, while Apple is the big wild card.
“there's only a few companies that have that.”