Worth a Look — Belridge Partners

Health Tech September 4, 2026

How Moderna Is Building a Personalized Cancer Vaccine for Every Patient

"The combination cut recurrence risk by 49% compared to Keytruda alone, which is already considered the best drug available for this cancer." — Stéphane Bancel, CEO Moderna

"More than a thousand cancer vaccine trials failed before this one, and most researchers had quietly given up on the idea that the immune system could be taught to fight a patient's specific tumor."

Stéphane Bancel, CEO · Moderna · a16z Podcast, September 2, 2026

01More than a thousand cancer vaccine trials failed before this one, and most researchers had quietly given up on the idea that the immune system could be taught to fight a patient's specific tumor.
02The Phase 3 trial enrolled 1,137 patients with high-risk melanoma who had already undergone surgery, which is exactly the setting where preventing recurrence matters most.
03Moderna builds a completely different vaccine for every single patient, which means no two doses that have ever been manufactured are the same medicine.
04Each vaccine encodes up to 34 mutations pulled directly from that patient's tumor, chosen because they are the ones the immune system is most likely to recognize and attack.
05The combination cut recurrence risk by 49% and distant metastasis risk by 59% compared to Keytruda alone, which is already considered the best drug available for this cancer.
06Keytruda has never been beaten in this setting before, and the fact that something outperformed it in a randomized Phase 3 trial is the reason the entire oncology field took notice.
07From the moment a surgeon removes the tumor to the day a patient receives their personalized dose, Moderna has compressed the entire manufacturing process to roughly 42 days.
08The stock moved 177% in a single session, the largest one-day percentage gain in the S&P 500 in 25 years, adding roughly $38 billion in market cap before the close.
09Moderna is now running studies in lung, kidney, bladder, pancreatic and gastric cancers, which together represent a far larger patient population than melanoma ever could.
10The same mRNA platform that was originally built for infectious disease is now being tested against autoimmune conditions, which suggests the technology has a much broader surface area than most people currently price in.

AI August 26, 2026

What Happens When the AI Boom Runs Out of Money

"Consumers do not want to pay for software. And consumers do not care about being productive." — Ben Thompson, Stratechery

"Consumers do not want to pay for software. And consumers do not care about being productive."

Ben Thompson · Stratechery · Invest Like the Best, August 18, 2026

01The US winning the AI race outright would be the most dangerous possible outcome because it gives China every incentive to destroy TSMC before accepting military inferiority.
02Chinese open-source models are not free. Kimi inference costs significantly more per answer than most people running it on their stack are accounting for.
03America wins by being more open, not by trying to copy China's top-down industrial policy.
04The AI buildout has moved from free cash flow, to debt markets, to Google issuing equity, to Nvidia tapping pension funds. The question is what comes after that if revenue doesn't arrive in time.
05The railroad parallel is not about infrastructure. It's about duration mismatch: the world ran out of money before the railroads made money, and the railroads still built the American economy.
06Google search is the most perfect aggregator ever built: zero marginal cost, scales in every direction, requires no investment to maintain. That is exactly why it cannot fund the AI buildout alone.
07The next shift for Meta advertising is from embedding-based matching to LLM prediction: not matching an ad to a person's profile, but predicting what that person will want to see before they know it.
08Apple is more insulated from AI disruption than any other big tech company because its core business is physical goods, retail, and distribution — none of which a language model can replicate.
09Microsoft's usage-based pricing for heavy AI users is dangerous because it makes enterprises think about what they are actually paying for, which is exactly the wrong question for Microsoft to have them asking.
10Amazon's first-best-customer model is the most durable compounding structure in technology: build the capability for your own business, then sell it to the world once it is proven.
11Every equity backstop Nvidia provides to a neo-cloud represents risk Nvidia is absorbing. Risk has a price. That price is a discount to Nvidia's true margin, not the reported one.
12Memory makers created such a large target on their own back through price manipulation that every major customer is now working to eliminate that dependency.
13If the AI bubble bursts, the lasting legacy will be power infrastructure, exactly as fiber in the ground was the lasting legacy of the dot-com collapse and railroad track was the legacy of the 1870s.

Robotics August 2026

The ChatGPT Moment for Robotics Is 12 Months Away

"In the US there will be 6 to 7 million fewer human laborers by 2030. There is only one thing possible to fill that gap." — David Reger, Neura Robotics

"In the US there will be 6 to 7 million fewer human laborers by 2030. In Europe, 7 to 10 million. In China, more than 80 million. There is only one thing possible to fill that gap."

David Reger, Founder & CEO · Neura Robotics · Machina Summit 2026 / theCUBE

01The ChatGPT moment for robotics is roughly 12 months away as embodied AI data pipelines finally come online.
02The core reason robotics lagged behind LLMs is that the training data did not exist: it has to be generated through physical embodiment.
03Physical AI requires a nervous system running at millisecond speed alongside a brain, not just a vision language action model.
04Labor shortages are the forcing function: 6–7 million fewer workers in the US by 2030, 80 million in China.
05Neura Robotics is building the full hardware stack not to own it permanently, but to define the standard and open source it.
06The platform play sits above the robot: an orchestration layer that gives robots situational awareness before they even move.
07Reger's core thesis is that companies should own their own robot intelligence the way they own their data, not outsource it to one centralized brain.
08Neura is partnering with Nvidia, Qualcomm, and Amazon to define compute and sensor standards rather than building everything alone.
09The business model is a "brain account": companies train their own models on Neura's platform, own their knowledge, and can license it to others.
10Open source wins in physical AI the same way it won in cloud: define the standard, remove the bottleneck, let the ecosystem scale.

AI August 2026

70% of Neolabs Will Die

"My guess would be that in three years we will have restrictions around access to Chinese open models." — Anastasios Angelopoulos, Arena

"My guess would be that in three years we will have restrictions around access to Chinese open models."

Anastasios Angelopoulos, Founder & CEO · Arena (LMSYS) · 20VC

01Two-thirds of the 75-plus neolabs will be acquired for parts or worth nothing.
02Kimi K3 beating American models proves China is doing more than just distilling US models.
03Enterprises want AI sovereignty: owning their full model stack without third-party supply chain risk.
04When Anthropic goes public, transparent margins will force inference price negotiations down industry-wide.
05A neolab at a $10B valuation needs roughly $4B in revenue within 2–3 years to justify it.
06The "worst case acqui-hire" logic makes the next fundraise nearly impossible for most neolabs.
07OpenRouter metrics overstate Chinese open source adoption because they exclude proprietary API usage.
08The US will likely restrict Chinese open source models within three years, driven by lobbying from frontier labs.
09Hosting a Chinese model locally does not eliminate backdoor risk, because attack vectors can be baked into training.
10Arena is seeing fake AI-generated job candidates pass real technical interviews, signaling a new wave of AI-powered corporate espionage.

Investing August 2026

The Secret to Beating 99% of Mutual Funds: Just Don't Sell

"We invested $400 million in Tesla between 2014 and 2016 and made $7.7 billion in profit." — Ron Baron, Baron Capital

"We invested $400 million in Tesla between 2014 and 2016 and made $7.7 billion in profit, and I think we're going to make five times that in the next 10 to 15 years."

Ron Baron, Founder · Baron Capital · The Compound and Friends, TCAF 252

01Started with $10 million under management in 1982, and as of June had $70 billion.
02Invested $400 million in Tesla between 2014 and 2016 and made $7.7 billion in profit; says they will make five times that in the next 10 to 15 years.
03Invested $2 billion in SpaceX across 27 transactions since 2017, and as of June it's worth about $25 billion.
04Buy and hold: the buy is the easy part. The hold is the hard part.
05Personal book value was $100,000 in 1982. It's now $4.5 billion.
06In 20 years, will robotics make higher-quality meals than humans? Yes. Will we have autonomous burritos? Yes. The question is just when.
07The price of a stock only tells you what you can buy and sell it for today. It doesn't tell you if a company is doing well or not.
08We won't sell because of timing. We sell when the competitive advantage deteriorates.
09About 11 to 12% of the money Baron manages is their own capital.
10Inflation is the hidden tax on every saver: the value of money falls by half roughly every 15 years at 4–5% inflation, meaning you must double your wealth every 15 years just to stay even.

AI Infrastructure August 2026

Why Even 5-Year-Old GPU Chips Are Getting More Expensive

"The A100 chip rental rate is going up and has stayed up and has not come down at all." — Steve Hou, Silicon Data

"The A100 chip rental rate is going up and has stayed up and has not come down at all, which tells you just how robust inference demand is."

Steve Hou, Head of Research · Silicon Data · Forward Guidance Podcast

01Silicon Data is building futures contracts for physical AI compute, letting buyers and sellers hedge GPU capacity risk like any commodity market.
02The token index is an expenditure-weighted price index, closer to PCE inflation than a raw demand counter.
03The index skews toward price-sensitive developers, making it a leading indicator of broader enterprise behavior.
04Token maxing drove the index higher, then plateaued as enterprises substituted toward cheaper models.
05The plateau is substitution, not demand collapse: users routed to cheaper alternatives, not away from AI.
06The A100, a five-year-old chip, has seen its rental rate hold and rise — the clearest signal of how strong inference demand actually is.
07The GPU forward curve shifted from backwardation to near-contango: cloud providers no longer discount long-term contracts and expect to raise prices on rollovers.
08At the one-year contract term, GPU rental prices rose monotonically even while short-term headlines signaled demand concerns.
09H100 spot softness while A100 and B200 prices rise is consistent with inference booming while training shifts to newer chips.
10Memory prices are rising because longer context and multimodal AI generate far more data than supply can absorb.

Defense Tech August 2026

The Era of the $15 Billion Aircraft Carrier Is Over

"The era of putting 5,000 people on a $15 billion aircraft carrier and using that for force projection is over." — Trae Stephens, Anduril

"The era of putting 5,000 people on a $15 billion aircraft carrier and using that for force projection is over."

Trae Stephens, Co-Founder · Anduril · Uncapped, Episode 35 with Jack Altman

01The traditional defense procurement model favors incumbents so heavily that even a perfect product from a new entrant gets rebid out to Lockheed Martin.
02Anduril's core thesis is that the defense industry needed a software-defined, hardware-enabled prime, and no such company existed when Founders Fund went looking.
03SpaceX and Palantir were both founded by billionaires, which made Trae Stephens skeptical a new defense prime could exist at all until his team pushed him to just build it.
04Navigating government procurement is so difficult that product is only 30% of the battle for a defense tech startup.
05The U.S. defense budget is bipartisan in a way virtually nothing else in Washington is, with the NDAA passing by supermajority every year.
06AI and autonomy policy actually originated inside the defense community, which has been studying and legislating autonomous systems for decades longer than Silicon Valley has.
07The close-in weapons system Phalanx (CIWS) has operated with full autonomy, identifying and engaging aerial threats without human targeting, for decades already.
08Anduril is building Arsenal-1, a 5 million square foot manufacturing campus, with 800,000 square feet coming online in early 2026, to shift from prototype to mass production.
09The manufacturing challenge at scale is entirely different from engineering a prototype: the question becomes how you mold exterior shells, make solid rocket motors, and source seeker materials at volume.
10A Patriot missile costs $2 to $4 million per unit, and Anduril believes new manufacturing approaches can produce comparable or superior capability at a fraction of that cost.

AI August 2026

95% of Your Token Spend Is on the Wrong Model

"95% of Your Token Spend Is on the Wrong Model." — Naveen Rao

"95% of Your Token Spend Is on the Wrong Model."

Naveen Rao · This Week in AI, Episode 16: AI Layoffs, Compute Costs & Agents

01Local models running at near-zero cost unlock 24/7 autonomous agents that per-token API pricing makes unaffordable.
02Open source at roughly Sonnet 4.5 capability handles background tasks without touching enterprise token budgets.
03Falling costs are forcing companies to finally match model quality to task complexity instead of defaulting to the most expensive option.
04Cheap tokens make the solo founder plus agents model viable for the first time.
05Unlimited local intelligence enables continuous monitoring and opportunity surfacing that frontier model costs would never justify.
06Open source being nearly as capable as frontier models is directly threatening the revenue model of the labs charging premium prices.
07Every time token costs drop, leaderboards and gamified metrics recreate the token maxing problem at a bigger scale.
08Spending more on tokens than on people is not sustainable because the productivity output still does not justify the cost for most companies.
09Cheap tokens did not solve the routing problem: people still default to the most expensive model because task complexity is hard to judge upfront.
10Intelligence per dollar still does not add up for the majority of enterprises regardless of where token prices sit today.

Investing August 2026

Bill Ackman on AI and the Frontier Model Risk

"I worry about the business model of the leading-edge models." — Bill Ackman

"I worry about the business model of the leading-edge models."

Bill Ackman · The Western Spirit Podcast with Ariel Litman

01Compute demand is structurally near infinite because intelligence is something individuals and enterprises are always willing to pay more for.
02The hyperscaler layer — Microsoft, Amazon, Meta, SpaceX — is the safest bet in AI because payback periods on data center investment remain extremely high even at current scale.
03The frontier model business model is under threat: open source models from China and others are now nearly as capable as paid frontier models for the vast majority of use cases.
04If a free or ad-supported model can accomplish what a $20/month frontier model does, the addressable market for paid frontier subscriptions shrinks dramatically.
05Anthropic reaching billion-dollar profitable quarters is a meaningful signal: it separates it from the narrative of endless frontier model losses requiring constant capital raises.
06The token budget crisis is real: companies are burning through annual AI budgets in a single quarter because employees have no incentive to use cheaper or more efficient models.
07Employees routing low-value queries through frontier models — like restaurant recommendations — is the clearest sign that enterprise AI spend is wildly inefficient and will be forced to rationalize.
08Microsoft is trading near all-time low earnings multiples despite being one of the primary infrastructure beneficiaries of the AI buildout.
09Amazon is in the same position: revenue and profit growth are accelerating while the stock's earnings multiple is near historical lows.
10The S&P 500 deserves a higher structural PE multiple than historical averages because the composition has shifted toward fast-growing, capital-light technology businesses.

AI Infrastructure August 2026

The Material at the Intersection of AI, Defense, and Fusion Energy

"You need tungsten nodes to not melt these high-end chips." — Geoffrey Woo, Antifund

"You need tungsten nodes to not melt these high-end chips."

Geoffrey Woo, Managing Partner · Antifund · The Pomp Podcast

01The CPU catch-up cycle is real: orchestrating thousands of parallel agent clusters shifts compute ratios back toward 1:1 GPU to CPU.
02When three unconnected smart people converge on the same bottleneck without coordinating, Woo's rule is simple: buy first, research second.
03Tungsten nodes are the thermal management layer inside high-end GPU chips, preventing meltdown at extreme compute loads.
04China controls the world's highest-purity tungsten supply and began restricting exports in January 2026.
05Japan's two largest WF6 producers permanently shut down as of July 1, 2026, removing 25% of global semiconductor-grade supply.
06WF6 prices are up over 200% year-over-year as a direct result of the Chinese export restriction.
07Tungsten hexafluoride is the chemical vapor deposited inside chips at 7nm and below to form the contact plugs connecting transistors.
08Sam Altman flew to Korea personally and secured roughly 40% of global memory supply before the trade became public.
09Fusion reactors require tungsten-based containment materials to survive plasma temperatures, creating a second demand spike independent of chips.
10The memory trade is now fairly priced. Tungsten has the same supply concentration and geopolitical exposure but has not yet been fully discovered by markets.

Defense Tech August 2026

AI Warfare and the Race for a Million-Drone Army

"Every country will have a million-drone army within 5 years." — Brandon Tseng, Shield AI

"Every country will have a million-drone army within 5 years."

Brandon Tseng, Co-Founder · Shield AI · Masters of Scale: Rapid Response

01Shield AI raised $2B at a $12.7B valuation in June 2026 and acquired Aechelon, a simulation company, to accelerate the AI training pipeline for autonomous systems at scale.
02Autonomous systems will outnumber human beings on Earth within this century.
03Every modern military will announce a million-drone army within 5 years.
04The cost curve has collapsed: VBAT does the same mission as a $40M Predator drone for $1M, and the US has lost 40 Predator and Reaper drones since October 2023.
05The pace of AI development is compressing build cycles from 150 engineers over 18 months to 2 engineers in 2 weeks for an AI fighter pilot.
06Defense is the proving ground for all physical AI: the same autonomy stack powering military drones will power self-driving cars, humanoid robots, and autonomous aircraft.
07The carrier strike group, America's primary conventional deterrent for 75 years, is now vulnerable to cheap long-range anti-ship missiles. Drone swarms are the replacement.
08Shield AI policy, US Department of Defense policy, and NATO policy all require a human in the kill chain, regardless of how autonomous the execution layer becomes.
09China's rate of innovation in autonomous military systems is accelerating faster than the US in classified data points Tseng has seen, even though the US currently leads.
10Two thirds of Shield AI's revenue is international, including Taiwan, India, and the Netherlands, making it a global autonomous defense platform, not just a US defense contractor.

AI Infrastructure August 2026

CPUs Are the Infrastructure Bet Hiding in Plain Sight

"For every $100,000 spent on GPUs, you are spending $5,000 on CPUs. That ratio is correcting now." — Dylan Patel

"For every $100,000 spent on GPUs, you are spending $5,000 on CPUs. That ratio is correcting now."

Dylan Patel · The Next Big Thing: A Megatrends Podcast

01CPUs, which were largely absent from AI server builds for three years, are now in aggressive catch-up demand.
02Memory prices are heading toward 85–90% gross margins as demand doubles while supply capacity grows only 20–30% per year, making memory the most acute pricing bottleneck in the entire AI stack.
03Two structural shifts are driving CPU demand: reinforcement learning needs CPUs to run environment checks on every output, and agentic AI needs CPUs to execute tool calls, web searches, code compilation, and real-world interactions.
04Every line of AI-generated code that gets deployed lands on a CPU. GitHub commits are up multiple X year-over-year, meaning CPU workloads are compounding alongside model usage.
05The GPU-to-CPU ratio has been severely imbalanced for years. That backlog is now being right-sized, creating a concentrated window of demand before it normalizes to steady-state incremental growth.
06ARM, Intel, AMD, Amazon Graviton, and Nvidia Vera are all competing for CPU share in AI infrastructure. Nvidia is guiding $20 billion in standalone CPU revenue with Vera, a product line that did not exist two years ago.
07The CPU architecture debate is not settled: Nvidia's Vera has fewer but faster cores optimized for agentic stall scenarios, while AMD and Amazon have more cores at lower per-core speed, better for parallel workloads. Both will win in different use cases.
08AI model costs are falling at roughly 60x per year, meaning the infrastructure buildout is becoming dramatically more efficient even as raw spend rises.
09Token efficiency is the real cost lever in enterprise AI, not switching to cheaper models: better models do more in fewer tokens, reducing both cost and human time per workflow.
10DeepSeek V3 was 600x cheaper than GPT-4 in two years, illustrating that the cost of a fixed quality level of AI is collapsing at a pace no other technology has matched.

AI August 2026

What Are the Biggest Risks to OpenAI and Anthropic Today?

"90% or greater of use cases can now be fully handled by many different models, including open source models." — Arvind Jain, Glean

"90% or greater of use cases [LLMs] can now be fully handled by many different models, including open source models."

Arvind Jain, Co-Founder & CEO · Glean · 20VC: "Why OpenAI and Anthropic Won't Win the App Layer"

01Consumption-based pricing could be the thing that breaks Microsoft's bundling advantage, since enterprises will route to whatever delivers the best output per dollar.
02The open source drive right now is coming from cost. It's not open source versus closed source. The only question is: are they okay with a Chinese model or not.
03The majority of enterprise AI workloads will run on open source models within 3 years.
04If an AI agent runs your workflows and you don't own the learning it accumulates, you've handed control of your business to OpenAI or Anthropic.
05Anthropic's vertical packs in design, legal, and finance are "quite shallow" and are expanding the market rather than taking share from incumbents.
06There is a sharp power law in enterprise AI adoption: a small group spends $10,000–15,000 per month in tokens, while most employees spend $20, almost all of it on basic question answering.
07Arvind pushed back directly on the CEO consensus that AI shrinks teams: if a competitor keeps 1,000 people with the same AI tools, they produce a 10x better product and win.
08On OpenRouter, the top six models by traffic are Chinese, with Anthropic sitting seventh — a strategic gap the US has not yet solved.
09Many nations that planned to build sovereign models a year ago have stepped back after realizing the investment required, though restrictions on Anthropic access in Europe are re-igniting the conversation.
10Anthropic is no longer just a model company: its MCP ecosystem of skills, automations, and integrations makes it an application-level competitor that Glean competes with directly every day.

Robotics July 2, 2026

The $2/hr Robot vs. the $40/hr Worker

"The economic incentive for adoption becomes almost impossible to ignore." — Andrew Kang, Robo Strategy

"A robot doing the work of three humans at roughly $2 an hour versus $35 to $40 for a US worker. The economic incentive for adoption becomes almost impossible to ignore."

Andrew Kang, CEO & Co-Founder of Robo Strategy · The Compound and Friends, June 29, 2026

01Andrew Kang invested $1M in Figure AI in early 2024 when his entire VC network told him not to. He is now running a publicly listed robotics fund.
02The most exciting robotics innovation is happening in private markets, locked away from retail investors. Robo Strategy (BOT) was built specifically to give public market investors access to pre-IPO robotics companies.
03Private valuations anchor to the last round and do not update until the next financing. That gap creates alpha for investors who understand the business.
04Global humanoid and quadruped robot shipments grew 250% last year to 53,000 units.
05Goldman Sachs projects 250,000 shipped by 2030 and a $38B total addressable market by 2035 — a number they just revised up sixfold.
06The total physical labor market is $50 trillion annually. That is the addressable market for humanoid robots working 24/7, no health insurance, no 401k, no breaks, no turnover.
07Morgan Stanley is forecasting 54% compound annual growth rate for humanoid revenue over the next decade. Jensen Huang has called it a $40 trillion market.
08Kang says intelligence gets solved in 1 to 2 years. The 3 to 4 year timeline is about building factories capable of producing millions of robots per year.
09The leasing model is where real commercial adoption happens first. Selling a six-figure robot is one conversation. Leasing one that pays for itself in months is a completely different conversation.
10Most robotics companies are moving toward lease-first.

AI Infrastructure July 2, 2026

Where Is the Moat in Foundational Models?

"The moat is not the model. It never was." — Marc Andreessen

"The moat is not the model. It never was."

Marc Andreessen · The a16z Show, January 7, 2026

01The model itself is no longer where value lives. Once a capability is proven possible, replication happens fast.
02Technological moats that used to last a decade now last a quarter. Speed of replication has permanently changed the economics of being first in AI.
03Andreessen's formula: product, integration, distribution, and captured value. None of those four words is "model."
04Distribution is the moat most people undervalue. Google did not win search because it had the best algorithm. It won because it became the default.
05In AI, whoever controls the default prompt interface controls the margin.
06Proprietary data is the only moat that does not commoditize. A competitor can replicate your model. They cannot replicate two years of your customers' workflow data fed back into your system.
07There is a well-defined intellectual pyramid in AI: god models at the top, small specialized models extending to embedded systems.
08Most enterprise value will be built not at the top of that pyramid but at the edges, where proprietary context compounds.
09In consumer AI, momentum is the moat. The model that gets used daily builds context, memory, and stickiness. The best model on a benchmark that nobody opens every day loses.

AI Economics July 2, 2026

Will Token Prices Fall 90%?

"The long-term token pricing should be 1/10 of what it is today." — Nikesh Arora, Palo Alto Networks

"I think the long-term token pricing should be 1/10 of what it is today. When that happens, you will see that people will consume more."

Nikesh Arora, CEO Palo Alto Networks · 20VC, June 22, 2026

01Token prices today are 10x where they should ultimately land. Nikesh expects dramatic reductions over the next 3–5 years as compute supply catches up and frontier models build sustainable business models.
02Free consumer AI is loss-making and consumes half of all global compute, pushing the cost burden onto enterprise and coding workloads that actually pay.
03Enterprise token budgets are a "use judiciously" model at Palo Alto — they monitor usage and only constrain employees who go off the rails, not the power users.
04The smartest employees will use 20x the tokens of average employees. Token-capping to control costs will harm your best AI talent first.
05Token spend as a % of developer salary is the wrong lens. The more interesting question: when tokens get cheap enough, does that number move from 3.8% to 15% or parity with salary?
06Frontier model companies are currently "value maxing, not token maxing" — at $1T+ valuations they need to show gross margin progress, so they're charging what the market bears.
07Ad-funded AI won't solve the consumer profitability problem. The total global advertising pie hasn't grown meaningfully in years, and 60–70% of it is already online.
08Physical AI is a pure depth play with no consumer use case. Models for planes, cars, and robotics will be separate, highly specialized, and deeply proprietary — potentially the most durable moats in all of AI.

AI June 25, 2026

Why Perplexity May Be Undervalued at $20B

"If Dario produces a better model, it's great for us. We benefit from every person's progress at every layer of the stack." — Aravind Srinivas

"If Jensen produces a better chip, it's great for us. If Dario produces a better model, it's great for us. If Apple produces a better device, it's great for us. We benefit from every person's progress at every layer of the stack."

Aravind Srinivas, CEO Perplexity · 20VC, 2026

01Perplexity is the only major AI product that orchestrates across competing models — GPT-5 and Claude Opus run inside the same product harness. OpenAI and Anthropic cannot say that about each other.
02Perplexity Computer is an orchestration system: a model paired with an agent harness that routes each task to the right model, tool, sub-agent, or local chip.
03Without a world-class harness, even a frontier model produces no business value.
04Every time any part of the AI stack improves — chips, models, harnesses — Perplexity's product gets better without Perplexity writing a line of model code.
05Perplexity's revenue tripled since early 2026, driven in part by Anthropic's model improvements. A competitor's R&D spend directly improved Perplexity's product.
06As enterprises fine-tune open-source models for specific domains, Perplexity can route to cheaper specialized models and compress inference costs further.
07"The model is no longer the product" (Greg Brockman). The orchestration layer above models is where economic value concentrates as models commoditize.
08Perplexity Computer's hybrid local-cloud architecture routes simple tasks to on-device models (zero token cost) and complex tasks to frontier cloud models.
09The $20B valuation was set in September 2025 on older revenue metrics. Revenue is now "far far more than before, growing really fast." IPO targeted sooner than 2028.

Robotics June 25, 2026

The ChatGPT Moment for Robotics Has a Higher Bar

"There is no internet you can scrape for the physical world." — Peter Chen, Covariant

"There is no data set to build a robotics foundation model. There is no internet you can scrape for the physical world. In order to have data, you have to build autonomously working systems for real customers."

Peter Chen, CEO Covariant · No Priors Ep. 48

01The ChatGPT moment for robotics has a higher bar than language: you need generality AND reliability, not 70% success with 30% catastrophic failure.
0299% of robots deployed in the world today are pre-programmed dumb machines — they repeat the same motion and cannot adapt to anything new.
03Covariant's foundation model is deployed across 10+ countries on 3 continents, all feeding one central brain.
04A single warehouse can hold 100,000 distinct items a robot must identify, grasp, and route without pre-programming.
05Covariant was founded in 2017 when no AI was good enough to make robots commercially useful.
06Every robot on Covariant's platform shares one model — everything each robot learns in the field improves all the others.
07The data flywheel mirrors Tesla: ship good-enough autonomy, collect real-world data at scale, build progressively smarter models.
08Simulation breaks down for manipulation because contact physics is too complex to model accurately.
09Language models need next-token prediction accuracy; robots need sub-millimeter spatial precision — a fundamentally harder grounding problem.
10The $500B+ global logistics and warehousing market is the first beachhead, driven by e-commerce growth and demographic collapse in warehouse labor supply.

Robotics June 18, 2026

One Brain, Any Robot: The $14B Bet on Skild AI

"There is no internet of robot data." — Deepak Pathak, Skild AI

"Robotics is a data problem. Unlike language or vision, there is not much data in robotics. There is no internet of robot data."

Deepak Pathak, Co-Founder Skild AI · NVIDIA AI Podcast Ep. 295

01There is no internet of robot data — the core reason physical AI is a harder problem than building LLMs.
02Skild raised $1.4 billion to build one brain that runs on any robot, any task, any form factor.
03$14 billion valuation for a company most investors have never heard of.
04Skild's omni-brain runs on humanoids, quadrupeds, robotic arms, and mobile robots from a single shared model.
0590% performance is easy in robotics. The last 10% is why robots have never gone mainstream.
06The digital world is less than 50 years old. Physical intelligence has existed for millions of years.
07The same pre-training and post-training separation that powers ChatGPT is now being applied to robots.
08Simulation can generate trillions of training examples per day but cannot fully replicate the real world.
09A classical factory robot requires a cage, precise measurements, and a setup that costs several times more than the robot itself.
10Edge compute on the robot itself is required because a robot that is falling cannot wait for a server response.