On capital and statecraft, written by an investor working at the intersection of the two | Building a multi-hyphenate life
🌍
Joined June 2021
- Tweets157
- Following2K
- Followers164
- Likes2.2K
Garima Sehra retweeted
Why open-weight models that compress margins at the model layer are net positive for AI infra demand, exhibit 1001.
An open-weight token consumes *roughly* the same amount of compute as a frontier token for a similar size model.
Garima Sehra retweeted
🛢️ The Gulf's Hormuz bypasses can carry only about half its oil
Before the war, about 20M b/d went through Hormuz.
Every Gulf producer is now trying to build a way around it.
🇸🇦 Saudi Arabia: the East-West pipeline can move 7M b/d to Yanbu on the Red Sea, but the ports there can only load about 4.5 to 5M b/d. Riyadh plans to expand them.
🇦🇪 UAE: the Habshan–Fujairah line carries about 1.5M b/d. A 2nd line, due in 2027, will double that and a 3rd is under study.
🇮🇶 Iraq: a dormant pipeline to Syria could be revived, and a Basra–Haditha line could branch to Syria, 🇹🇷Türkiye and 🇯🇴Jordan. The Jordan route alone is estimated at about $10B for roughly 1M b/d.
Even if every project is built, total bypass capacity comes to roughly 9 to 10M b/d.
That's about half of what used to pass through the strait.
🇰🇼Kuwait, 🇧🇭Bahrain and 🇶🇦Qatar have no bypass at all, and Qatar's LNG can't be piped anywhere...
Every bypass brings its own chokepoint.
Saudi oil diverted to Yanbu still has to leave the Red Sea, where the Houthis claimed an attack on Aramco's Rabigh refinery this week and where a drone shut the East-West pipeline last month.
Iraq's routes run through Anbar and Syria.
Fujairah sits on the Gulf of Oman, just outside the strait.
The pipelines spread the risk around.
They don't remove it, and they cost billions to buy that relief.
The uncertainty is timing... The UAE line is the only one with a firm date and Iraq's plans are mostly studies.
If Hormuz reopens first, the urgency, and the funding, could fade.
Source: Bloomberg
Garima Sehra retweeted
Before he co-founded Anthropic, Dario Amodei spent years recording neurons in salamander retinas.
His PhD advisor, Michael Berry, said he was the most talented graduate student he ever had.
His neuroscience work is rarely discussed, but it explains a lot about how he thinks.
He started graduate school at Princeton in theoretical physics. In 2006, his father, Riccardo, died of hepatitis C. Dario moved his graduate work to biology to work on human disease. A few years later, a new treatment made hepatitis C about 95% curable.
Within a few years of his father's death, the cure rate for that disease went from about 50% to about 95%.
In 2007 he became a Hertz Fellow and joined Berry's retina lab. The retina is part of the central nervous system, and a salamander's can be kept alive in a dish and shown movies.
The tools for measuring the retina weren't good enough for him, so he co-invented a better sensor, which was unusual for the lab.
In his words, "my time in neuroscience was spent building probes to study small groups of neurons."
Berry has said Dario was the most talented graduate student he ever had.
He didn't care what a single neuron did. He wanted to know what a whole population did together.
He called his thesis "Network-Scale Electrophysiology: Measuring and Understanding the Collective Behavior of Neural Circuits." It won the Hertz Thesis Prize.
In 2012 the work made the cover of the Journal of Neuroscience.
Dario and his co-first author recorded more than 200 ganglion cells in a salamander retina at once, as many as 95% of the cells over the electrode array.
The paper described it as "unprecedented access to the complete neural representation of visual information."
Three years later, he co-authored a paper in PNAS showing that the collective activity of up to 160 retinal neurons was "poised near an unusual kind of critical point."
After Princeton, he went to the Stanford School of Medicine as a postdoc.
In Parag Mallick's lab, he moved from neurons to proteins, working on mass spectrometry and proteomics, and co-authored a Nature Biotechnology paper on an open-source toolkit for mass spectrometry data.
From there he went to Baidu, then Google Brain, then OpenAI, and finally Anthropic.
In his 2024 essay Machines of Loving Grace, he wrote that when he was in neuroscience, many researchers were asking the wrong questions about learning because the idea of scaling didn't exist yet.
To him, scaling was a high-level explanation of how intelligence works and why it evolved so easily.
Biological and artificial neurons work differently, he argued, but they pose the same question, which is how a trained network of simple units computes.
He pointed out that a mechanism found inside AI models was later found again in the brains of mice.
And he predicted AI could deliver the next century of neuroscience in five to ten years.
Andrew Huberman on why Meta, OpenAI, and Anthropic are all going to become biotech companies.
"Dario came up through neuroscience. Sam, I've spoken to, is extremely smart about the neuroscience, as is Zuck.
The technologists who used to just engineer stuff, they're becoming neuroscientists...and I think most people don't realize that.
They will all tell you, 'Okay, here's the goal. The goal is to be able to read and write from your brain non-invasively.'
Even just for things like drive and motivation, a non-invasive device where you could literally crank up your level of motivation by stimulating the nucleus accumbens and related pathways while engaging in a particular learning task or in hard work, and then turning it off so that you can capture more energy.
That's coming, and those companies are all about that. That's what the real arms race of AI is about."
Garima Sehra retweeted
🛢️🏭 Refining Capacity Is Strategic Power
Crude oil only matters once it becomes something usable.
-Gasoline
-Diesel
-Jet fuel
-LPG
-Petrochemicals
None of it exists until crude passes through a refinery.
This week showed what happens when that step gets squeezed.
US diesel hit $6.0556 a gallon, a new all time high, while Gulf Coast refiners run at their busiest since 2019.
The world's biggest refineries tell the same story at scale.
India's Jamnagar complex, run by Reliance Industries, processes 1.4 million barrels a day, the largest single refining site on earth, in a country that produces a fraction of the crude it refines.
Nigeria built the Dangote refinery, 650,000 barrels a day, to stop importing the fuel made from its own oil.
Saudi Arabia, the top crude exporter, still built over 3.3 million barrels a day of refining capacity, because raw crude leaves the margin on the table.
Crude reserves make headlines.
Refining capacity is what turns a barrel into a truck moving or a plane taking off, and this week showed exactly what happens to prices when that capacity runs out of room.
themerchantsnews.substack.co…
Garima Sehra retweeted
Humanoid robots will become one of the biggest industrial markets of the next decade.
The chart estimates that the global humanoid robot TAM could grow from approximately $613 million in 2025 to $138 billion by 2035.
That would represent an extraordinary expansion as robots move from prototypes and pilot programs into factories, warehouses, logistics centers, and other commercial environments.
The forecast also shows annual shipments increasing from approximately 14,700 robots in 2025 to nearly 6.5 million robots by 2035.
The economics could also become increasingly attractive as production scales and manufacturing improves.
The chart estimates that the average selling price could decline from approximately $41,800 in 2025 to about $21,300 by 2035.
At the same time, the bill of materials could fall from approximately $27,700 to roughly $13,500, creating more room for manufacturers and suppliers to earn attractive margins.
This cost reduction is important because robots do not need to replace every human worker to create demand.
They only need to become economically competitive in industries facing labor shortages, rising wages, unsafe working conditions, or high employee turnover.
Here is how you can benefit from all of this.
The supply chain could be one of the largest beneficiaries because every humanoid robot requires processors, memory, cameras, LiDAR, sensors, actuators, precision gears, batteries and power management components.
NVIDIA, Intel, and Qualcomm could benefit from supplying the computing hardware, while companies such as Harmonic Drive, Nidec, Regal Rexnord, Hyundai Mobis, and Inovance could benefit from supplying motors, actuators, and motion control systems.
CATL, LG Energy Solution, Samsung SDI, SK On, and Panasonic could benefit from the battery demand created by large robot fleets.
Companies such as Sony, Murata, Novanta, Hesai, RoboSense, Ouster, and LG Innotek could benefit from supplying cameras, sensors, LiDAR, and other perception technologies.
The opportunity will extend beyond the initial sale because humanoid robots will require software updates, replacement parts, maintenance, fleet management, charging infrastructure, and ongoing artificial intelligence services.
Goldman Sachs previously estimated a $38 billion humanoid robot market by 2035, while Morgan Stanley has projected that the broader humanoid economy could eventually reach approximately $5 trillion by 2050.
The near term opportunity is likely to come from industrial automation, but the long term opportunity could expand into healthcare, retail, construction, hospitality, and household applications.
If you enjoyed reading this make sure to follow @MelvinInvests for more robotics and semiconductor insights, and if you want to see exactly what I'm buying as an analyst at Milk Road Pro, check out this link and come join us.
link.milkroad.com/tx5qip
The next memory supercycle is hiding inside humanoid robots (Save this).
This chart estimates that each humanoid robot could contain $600–$800 or more in memory related and semiconductor components, including DRAM, HBM, NAND flash, sensors, communication chips, and other devices.
If manufacturers eventually produce 1 million humanoid robots, that could represent approximately $600 million–$800 million in memory content, while 10 million robots could create a $6 billion–$8 billion opportunity.
However, the larger opportunity may come from the data centers that train these robots and process the enormous amount of information they generate.
AI servers already require significant amounts of HBM, DRAM and enterprise SSD storage and HBM demand continues to rise as companies build larger systems for training and running AI models.
Humanoid robots could create an additional source of memory demand on top of the existing artificial intelligence infrastructure cycle.
Every robot needs memory to process sensor data, recognize objects, understand its surroundings, and execute complex movements in real time.
As robots become more autonomous, they may need to process more information locally instead of sending every task to the cloud, which could increase the amount of high performance memory required inside each machine.
Humanoid robots could also generate enormous amounts of video, sensor, and operational data that must be stored, analyzed, and used to improve future generations of robots.
This would create additional demand for NAND flash and enterprise SSDs inside the data centers supporting these systems.
Enterprise SSD demand is already benefiting from AI workloads, while meaningful new memory capacity may take years to come online.
As we have seenAI creates the first major wave of memory demand, while humanoid robots extend that demand from data centers into the physical world.
HBM could benefit from model training and inference, DRAM could benefit from on device computing, and NAND flash could benefit from the storage requirements created by robot generated data.
And when humanoid robots become a mass market product, memory could become one of the largest beneficiaries inside the bill of materials.
If you enjoyed reading this make sure to follow @MelvinInvests for more semiconductor and robotics insights and if you want to see exactly what I'm buying as an analyst at Milk Road Pro, check out the link below.
Garima Sehra retweeted
The consumer agent just started to feel real.
I told Instinct: “You know that thing we did? Move it back 90 minutes.”
It remembered the booking, the pub’s website, changed the reservation and sent me the confirmation.
Most businesses are not ready.
Garima Sehra retweeted
What's the next country to blow up? Our Datacenter Model subscribers got the answer back in February: India. Scaling to just under 10GW by 2030, India enters the top 4 global datacenter markets! All the AI Labs and hyperscalers are itching for capacity. (1/3)🧵
Garima Sehra retweeted
We are witnessing the largest CapEx spending spree in history.
Hyperscalers have committed a record $2.6 trillion in future spending across data center leases and equipment purchases.
This includes leases for data centers and power infrastructure that have not yet commenced, and existing purchase contracts for equipment and services.
These obligations do not appear on company balance sheets and are disclosed only in limited accounting footnotes, despite their enormous size and duration.
Alphabet, $GOOGL, has the largest total, with $811 billion in purchase commitments and $85 billion in leases not yet commenced, bringing its combined obligations to ~$896 billion.
By contrast, Oracle, $ORCL, has committed $32 billion to purchases, but $260 billion to leases that have not yet commenced, the largest lease obligation among the group, bringing its total to ~$292 billion.
The AI buildout has become one of the largest long-term capital commitments in corporate history.
Garima Sehra retweeted
🚨Add a 4th chokepoint to the list, and this one has nothing to do with any war.
The Panama Canal Authority is suspending part of its booking system, daily auctions for Panamax class locks, starting July 25, because of water levels at Gatun Lake and the risk of a new El Niño.
Hormuz, Bab el-Mandeb, and the Black Sea are jammed by conflict.
Panama is jammed by weather.
Churchill's warning about never depending on 1 route did not specify which kind of disruption.
themerchantsnews.substack.co…
🚨113 years ago this week, Churchill warned that safety in OIL lies in variety, never in a single route.
The market spent June forgetting it.
Then this weekend 3 export straits jammed at once, from 3 unrelated wars.
Brent had already round-tripped from its June lows back toward the $ 90s.
Backwardation is back.
Most desks called June's Hormuz reopening the end of the war premium.
I think they misread the tape, and I lay out exactly why in my latest article.
Inside it I map the 3 chokepoints jamming in real time.
Why the fastest hedge fund oil build since 2016 is both a confirmation and a trap, the full scenario map and the bypass buildout quietly redrawing the map from Yanbu to Baniyas, with US majors as the counterparty.
Read it here 👇
open.substack.com/pub/themer…
Garima Sehra retweeted
Trump’s Pipeline Wars: How Iran’s Gambit Exposed China
War is the continuation of politics by other means and Trump has moved that logic from the battlefield to the barrel.
The pipeline wars are his answer to Iran’s closure of Hormuz: a counter‑offensive that punishes Tehran, exposes China, and pulls Iraq, Syria, and Venezuela toward the American orbit as emerging allies.
Iran’s gambit was revolutionary bravado: slam shut the strait, choke off a third of seaborne oil, and dare America to blink. Iraq’s exports, long 90‑plus percent dependent on Hormuz—collapsed, and Baghdad discovered it was less an energy state than a client of a narrow waterway patrolled by a hostile regime and an American carrier group.
The Strait has been war‑gamed for decades. Trump tried diplomacy first. Now the world is watching what hard power looks like when the simulations go live.
His answer is to treat Hormuz as a flaw, not fate. Epic Fury broke Iran’s ability to escalate; the strategic move is what follows—build around Iran. Push Iraqi barrels toward Turkey’s Ceyhan.
Revive Mediterranean outlets. Bring back the Kirkuk–Baniyas concept: an old 1950s line from Kirkuk to Syria’s port of Baniyas, shuttered by war and neglect, now reborn as a 300,000–700,000 barrel‑per‑day artery with U.S. backing, American firms doing the studies, and sanctions eased just enough to lay steel. Iraq gets cheaper exports and diversification away from both Hormuz and Ceyhan. A post‑Assad Syria stops being a crater and starts being a corridor, earning hundreds of millions in transit fees, plus jobs and infrastructure. In practice, Syria and Iraq are being bound into an American‑centric energy system, precisely how fragile states become durable allies.
The deeper casualty is China. Beijing built its industrial machine on discounted barrels from Iran and Venezuela, moved by shadow fleets through long, vulnerable sea lanes. That is not diversification; it is dependency. Churchill warned that “safety and certainty in oil lie in variety of supply.” China concentrated risk in sanctioned regimes and contested waters, then called it strategy.
Venezuela shows the Trump doctrine at work. For years, Caracas was a major producer, enabler of state‑sponsored terrorism, and a willing ally of China.
The takedown of Nicolás Maduro was a seismic event largely ignored in polite foreign‑policy circles: a hostile petro‑regime toppled without occupation, then flipped into a grudging supplier to U.S. refiners. Oil long routed to China now flows to America, strengthening U.S. energy security and starving Beijing of friendly heavy crude.
This is Hegel’s dialectic in hard assets. Iran’s closure of Hormuz is the thesis. Trump’s strikes, pipelines, and Venezuela turn are the antithesis. The pipeline wars are the synthesis: chokepoints contested, transit states turned into allies, hostile producers pulled into an American system.
In 2026, the map matters again and Trump is redrawing it so America sits on the pipes while Iran and China sit on exposure.
The United States welcomes the signing of a historic MOU between the Governments of Iraq and Syria on the rehabilitation and reconstruction of the Iraq-Syria crude oil pipeline. This priority infrastructure project will advance security and stability through regional interconnectivity and prosperity. U.S. companies will play a key role in advancing construction.
state.gov/releases/office-of…
Garima Sehra retweeted
BofA : Same AI data center, different value perception. You'd be surprised to know why.
Every mega-cap is pouring hundreds of billions into AI data centers. Is that capex creating value and why is META's AI capacity valued at a fraction of Google's and Amazon's?
BofA has issued its report answering the question. Better bookmark it
Same buildout, wildly different market pricing. BofA strips out each company's legacy business and back-calculates the market's implied value for their AI capacity.
The result: the market pays for capacity it can see monetized (cloud contracts), and discounts capacity it can't (META's internal-only use case).
Implied value per GW of AI capacity: $GOOG ~$110B/GW, Amazon ~$59B/GW, META ~$4B/GW. Meta trades at a ~27x discount to Alphabet on the same physical asset.
BofA raised 2027 capex forecasts: Alphabet $257B → $290B, $META $157B → $185B, AWS $196B → $230B (2026: Alphabet $187B→$195B, Meta $130B→$145B, AWS held at $159B). These are BofA model figures, above official guidance in places (Meta's public 2026 guide is $125–145B).
Cost per incremental GW (2026): $AMZN ~$25B, Google ~$37B, META ~$45B. META is both the smallest adder and the most expensive per GW (external GPUs + upfront civil works).
Why the gap: cloud revenue per GW (2026) - AWS ~$10.6B, Google Cloud ~$15.7B maps capacity to contracted revenue. META has no external cloud line; capacity feeds ad efficiency, not sellable compute.
Bull case: in BofA's optimistic scenario META reaches ~22.8–23GW by 2030; if 40% is sold as enterprise AI at ~$12B/GW, that's a ~$110B enterprise-revenue opportunity which is why the "Meta is cheap" narrative exists. Model assumption, not guidance.
Constraints flagged: power access (DOE/EPRI cited: data centers ~4% of US power in 2023 → up to ~9% by 2030), GPU/build delivery timing, and enterprise willingness to pay.
AI-capex debate isn't about who spends most. It's whether spend converts to sellable revenue. The market already prices that in: GOOG ~$110B/GW, AMZN ~$59B/GW, META ~$4B/GW.
META is the cheapest AI asset in Big Tech because it feeds ads, not a cloud P&L. This is about to change now as they are moving into cloud.
My Read: They didn't do this because they had "excess compute". They did it because they need the market to re-rate them.
Coincidentally META plans to double its compute capacity to 14GW in 2027 and begin manufacturing its in-house Iris AI chip in September.
Are you still surprised is moving into cloud business with their "excess compute"?
Repost this to let the world know the context behind META's decision to sell "excess compute"
Garima Sehra retweeted
BREAKING: China’s crude oil imports dropped -41% YoY in June, to 29 million tons or 7 million barrels per day, the lowest in 10 years.
MoM, oil imports fell by -4 million tons, or -12%, to the weakest level in 8 years.
Import volumes have now declined -26 million tons, or -47%, since the late-2025 peak.
By comparison, the average between 2020 and 2025 was ~45 million tons per month.
Meanwhile, China's seaborne crude imports stood at around 6 million barrels per day in June, with imports from the Middle East falling to their lowest in 10 years.
At the same time, Iranian oil imports fell -40% MoM, to below 800,000 barrels per day.
The Iran War has created a historic supply shock.
Garima Sehra retweeted
The mega bull case for AI infrastructure would be *if* market share shifted away from certain frontier labs with 90%+ inference margins toward cheaper models, whether open-source or closed.
It would increase the ROI on AI spend for end customers by increasing intelligence per dollar, which would drive incremental token demand. Margin dollars would effectively get redistributed from the frontier labs to AI infrastructure providers. The infra winners would be those with the lowest per token cost and the winners at the model layer would be those with the highest token efficiency.
There are many reasons Jensen is so focused on open source, but this is likely the most important one as I think he is probably less worried about a monopsony these days. Lower margin % at the model layer = more margin $ at the infra layer all else equal.
With SpaceX and Meta being vertically integrated and possessing the #3 and #4 models respectively it is more possible than ever. Note that Grok 4.5 is ahead of Fable for some useful tasks at a much lower cost, so ranking them #3 is conservative.
This is not happening yet. Cheap, mostly open source tokens are likely the majority of volume today but the majority of economic value is still accruing to the most intelligent models. Might change though.
We will see.
This is true as I have heard this from contacts in the Valley. Goes with my pinned post.
The AI race is shifting from bigger models to cheaper, smarter systems cnbc.com/2026/07/10/the-ai-r…
Garima Sehra retweeted
Most people think Harry Potter is the best selling book series in history at ~615M books.
That will likely change this year.
It will be overtaken by One Piece, by Eiichiro Oda, regarded as the greatest Japanese manga / graphic novel of all time, which just crossed 600M in March and is projected to surpass HP this year and hit 725M+ when it ends in ~2030-32 (there’s 10-15% left).
The story follows Monkey D Luffy, a young boy who has rubber man powers, who assembles a pirate crew and goes on an epic journey to find the legendary “One Piece” treasure.
As a voracious reader, I struggled to take manga seriously for most of my life. I’m a hardcore Harry Potter fan, and even thinking of manga in the same breath once disgusted me. I relegated manga to a “niche comic book for nerds” and anime to a “kids cartoon”. “These aren’t real books, it’s just for people who can’t read longform”, I thought. But the past few years have completely changed my mind. I’ve been personally reading and watching (YouTube/Crunchyroll) many manga / anime including One Piece. Aside from the appreciation for the artistry of the characters, the stories strike me as immensely powerful with deep lessons about loyalty, sacrifice, friendship, ambition, justice, power, grief and kindness. Once you get in the flow, it’s far more enjoyable than most non-fiction and rivals anything on my shelf. The storytelling is beautiful. It awakens the child inside you without being juvenile. I highly recommend it to anyone who has previously had reservations like I once did, and doubly so if you want to enjoy something with your kids.
Garima Sehra retweeted
Dylan Patel just mapped out the most important investment theme in AI infrastructure (Save this).
"In about two years, solar plus battery will be cheaper than gas."
Every new NVIDIA Blackwell rack pulls 120 kilowatts, Rubin Ultra rack pulls 600 kilowatts and the next generation hits a megawatt.
The US grid cannot keep up, interconnection queues now run five years in many markets so the entire industry is being forced to solve power from first principles.
The solar thesis is already happening.
BloombergNEF's 2026 LCOE report, covering 800+ financed projects across 50+ markets puts solar plus 4 hour battery storage at $57 per megawatt-hour.
Combined cycle gas turbines hit $102 per megawatt hour, the highest on record, up 16% year over year.
In California and parts of Texas, solar plus storage is already cheaper than gas for data center power today and solar panel costs are expected to drop another 30% by 2035.
Getting power from the grid into the form chips actually require is an entire industry unto itself and NVIDIA just rewrote the rules.
The 800 volt DC transition is the most important infrastructure shift that's happening right now.
Today's data centers run on 48 volt DC power delivery, a single next-generation GPU pulls over 2,500 watts and at 48 volts, the current required to power a megawatt rack would melt the copper wiring.
The investment thesis breaks into four layers and the first layer is power semiconductors, specifically silicon carbide and gallium nitride.
At 800 volts, traditional silicon based IGBTs hit their physical limits. SiC and GaN devices are the mandatory replacement.
Infineon estimates $175,000 of semiconductor content per megawatt of AI rack power, versus almost nothing today and by 2030, power semiconductor content per AI cabinet grows from $15,000 to $115,000+.
The names here are Infineon ($IFNNY), ON Semiconductor ($ON), Wolfspeed ($WOLF), Navitas ($NVTS), and STMicroelectronics ($STM).
The second layer is power management and conversion.
Vertiv ($VRT) is NVIDIA's lead architectural collaborator for the 800V transition, building the hardware that converts grid AC to 800V DC and the DC to DC power shelves for ultra dense racks.
Eaton ($ETN) and Monolithic Power Systems ($MPWR) round out this layer.
The third layer is grid to site infrastructure, GE Vernova ($GEV) builds the heavy electrical equipment that connects utility power to the data center campus.
Orders are running at twice the rate of shipments, the classic leading indicator of sustained multi year revenue growth.
The fourth layer is behind the meter power generation like your bloom energy because grid interconnection queues run five years, hyperscalers are bypassing the grid entirely, building dedicated gas, solar and battery systems on site.
Make sure to follow me @MelvinInvests for more opportunities across the AI supply chain.
Garima Sehra retweeted
19 million barrels🛢️
That's the entire cushion left in the US Strategic Petroleum Reserve.
SPR sits at about 319.5 million barrels, the lowest since 1983
Estimated minimum operating level: around 300 million barrels
The gap between the 2: roughly 19 million barrels before hitting the floor
Refilling it back to historical norms could take years and tens of billions of dollars
Built after the 1973 oil embargo so the US would never get caught flat footed again.
We're close to finding out what happens if it does.
Garima Sehra retweeted
Three years ago, two Harvard dropouts set out to build a better AI chip than the largest companies in the world.
Almost everyone I called at the time said it was impossible.
Today, Etched (@Etched) comes out of stealth with $800M total raised, $1B in signed customer contracts, and a working next-gen AI chip.
This was my excuse to ask the two founders, @UbertiGavin and @robertwachen, every question I have about compute and inference.
We discuss:
- Why they built an entire rack and not just a chip
- The two technical bets behind their architecture no one else has tried
- How two founders in their twenties recruited industry legends
- The night they nearly ran out of money
- Why whoever produces the most tokens wins
If you care about the future of compute, Gavin and Rob are two people to know. I think you will find the story of what they have built hard to forget.
Enjoy!
TIMESTAMPS
0:00 Intro
1:00 Why Nobody Believed Etched Would Work
14:06 Why Inference Is the Bottleneck
22:27 Gavin and Rob’s Origin Stories
33:24 Taking Huge Risks to Move Faster
49:43 Kernels, Compilers, and the AI Stack
1:02:08 Raising $100M to Survive
1:16:00 The Future of Models, Agents, and Intelligence