· Jonathan Ross

Jonathan Ross: Groq — Inference Speed, Return on Luck, and Intentional Leadership

In the AI age competitive advantage shifts from answering questions to asking the right ones; a founder's real job is full-time change management executed through minimal-constraint goals, intentional leadership, and hiring that selects against negatives — while seizing luck faster than rivals is what actually compounds into dominance.

aiinferenceleadershiphiringchange-managementfundraisingfounder-modesemiconductors0% confidence

Why this is in the corpus

Jonathan Ross created Google's TPU and founded Groq (LPU inference chips), closing a ~$20B Nvidia partnership in a three-week window. The interview is an unusually dense playbook spanning leadership-as-infinite-forms, single-metric alignment, no-one-on-ones politics control, Return on Luck, salary-for-equity survival financing, loss-bias hiring, and change management — a framework-and-play goldmine for operators.

Summary for skimmers

Groq founder on complementary GPU+LPU inference, the three-week Nvidia deal, asking-the-right-questions as the AI-age skill, leadership having infinite valid forms, the 25M-tokens/sec challenge coin, killing one-on-ones to kill politics, confidence via shadowing, Reality Quotient and dominant-game selection, change management as making change feel like no change, Return on Luck and three missed LLM opportunities, Grok Bonds salary-for-equity survival, hiring for loss bias / booking the win early, manufactured discontent, and Intentional Leadership.

Briefing

What survives the editorial filter

This page should feel like a smart colleague already listened for you and left only the operating logic worth keeping. Not everything said in the episode makes it through.

Trust signal

Direct episode extraction

Best used for

Decision-grade retrieval metadata not yet added for this episode.

Hold lightly

No explicit downgrade reason stored yet for this episode.

Principles

Durable claims that survive beyond the speaker's biography — each with explicit limits, transferability judgment, and evidence.

Principle

Build trust by selling customers only what they actually need

Customer-need simplicity beats clever positioning.

Ross contrasts his earlier tendency to play 3D chess with Jensens relentless what-does-the-customer-need focus: tell customers only things that are true and supportable, refuse to sell what they do not need, and trust follows.

If you would not buy it in their shoes, do not sell it to them.

“I got way too cute on trying to play 3D chess, whereas Jensen is very much just like, what does the customer need?”Jonathan Ross

“Im gonna sell things to customers that they actually need and that I believe they need”Jonathan Ross

Principle

The better the people, the harder they are to manage

Elite creative talent raises coordination cost, not lowers it.

Ross draws a military analogy: higher rank allows more reports, but managing scientists requires dramatically smaller spans. His 450-person creative org felt like managing 5,000 in coordination difficulty (though smaller in that innovations happened on their own).

Budget more management bandwidth per elite hire, not less.

“the better the people, the harder they are to manage”Jonathan Ross

“it probably was more like managing a group of 5,000 people than it was 450 in many ways”Jonathan Ross

Principle

Sustained performance runs on manufactured discontent

Deliberately stay dissatisfied to keep driving after success.

Ross observed that in a room of successful people, entrepreneurs were least happy with their wealth (and thus started more companies) while others were discontent with prior work product. Everyone successful had something they were discontent about. He cites Ogilvys divine discontent and Steve Jobs going straight to the next product.

Engineer a source of discontent; contentment is the enemy of continued innovation.

“a lot of the most successful entrepreneurs, they sort of manufacture their own sort of discontent”Jonathan Ross

“You have to have a personality where you are constantly discontent if youre gonna keep pushing things forward”Jonathan Ross

Principle

Thinking faster makes a model think smarter

Speed is a quality lever, not just a latency lever.

Ross explains via AlphaGos jump in ELO on TPUs and the discovery of Move 37, which GPUs at the time could not find because it was too deep in the search chain. Faster hardware lets the model reflect and reconsider second-best moves that turn out better in context.

Treat inference speed as an input to answer quality, not a nice-to-have.

“being able to think faster makes you think smarter”Jonathan Ross

“you put these LPs into a system and all of a sudden the generation of tokens gets faster and its like getting broadband instantly on these existing models”Jonathan Ross

Principle

A leader must hold confidence in a decision so others have confidence to execute

Displayed conviction, not decision content, unlocks team execution.

Ross realized after shadowing an experienced CEO that he had been making the right calls all along but lacked confidence; acting with confidence (without changing the decisions) changed how people followed. He notes low-confidence people think more deeply but must still learn to act decisively.

Separate the analysis (do it deeply) from the delivery (do it with conviction).

“I needed to have the confidence in a decision so that other people wouldve the confidence to execute”Jonathan Ross

“I didnt change my decisions, but it changed my leadership”Jonathan Ross

Principle

In the AI age, advantage shifts from answering questions to asking the right ones

When answers are free, the question is the leverage point.

Schooling trained people to memorize and answer; AI inverts this. Because the model can research and solve, the quality of the prompt-question determines the quality of the output. Ross ties this to everyone moving from individual contributor to leader-of-AI, where the leader's job is asking the question no one else asked.

Train yourself and your kids to ask better questions, not to store answers.

“success in the information age was about being able to answer questions. Success in the AI age will be about being able to Ask The Right Questions”Jonathan Ross

Principle

People innovate in domains where they hit the problem before everyone else

Early exposure to the problem is half of the innovation equation.

Ross uses his AlphaGo-on-TPU experience: he saw hardware transform model performance before others, which let him recognize the fast-inference opportunity. He argues you need both the early exposure and the willingness to seize on it.

Seek proximity to emerging problems; perception advantage compounds with seizing behavior.

“a lot of really good innovators are innovators because they experienced the problem before others did”Jonathan Ross

“The futures already here, its just not evenly distributed”Jonathan Ross

Principle

Fewer constraints give a team more freedom to surprise you with innovation

Under-specify the how so the team can out-innovate you.

Ross frames disruption as inherently requiring doing things differently; a team that can only execute your instructions cannot surprise you. But he warns this requires crisply distilling what actually matters, or you will accidentally over- or under-constrain.

Specify the goal precisely and the method loosely.

“The Fewer Constraints, yeah. That you give someone, the more freedom they have to solve the problem”Jonathan Ross

“The only way for your team to innovate, right, without you being the innovator, is they must be able to surprise you in a good way. Which means you must not over constrain the goal”Jonathan Ross

Principle

Reality Quotient beats IQ: the ability to recognize and choose the dominant game

Winning is choosing the right game, not out-playing the wrong one.

Ross distinguishes RQ from IQ: some brilliant people would not recognize reality if it tapped them on the shoulder. The exemplar is MySpace optimizing accounts-signed-up while Facebook optimized monthly active users; maximizing the dominant metric beats maximizing the conventional one.

Before optimizing, ask whether you are even measuring the metric that wins.

“reality quotient, which is different from intelligence quotient”Jonathan Ross

“in the most extreme form, its the ability to choose the dominant game thats being played”Jonathan Ross

Principle

Leadership has infinite valid forms; copy the one that is true to you

There is no single correct leadership style, only the one authentic to you.

Ross contrasts himself (an autonomous delegator who has no drivers license and delegates constantly) with the control-freak founders common on the show. Trying to be a control freak would have failed for him; a control-freak leader should double down on command-and-control. The failure mode is executing advice that is not true to you.

Diagnose your natural style first, then amplify it rather than importing a foreign playbook.

“they dont realize that there is an infinite number of ways to be a leader”Jonathan Ross

“you just have to pick the form of leadership that works for you”Jonathan Ross

Principle

Extreme performance comes from one brutally clear priority

One coin-sized metric aligns an entire creative org.

Senra invokes Kelly Johnson of Skunk Works; Ross operationalizes it with the 25M-tokens-per-second challenge coin. Chip, software, power-cost, supply-chain, and datacenter teams could all connect their work to the single number without being told how.

If you cannot fit the goal on a challenge coin, it is not yet clear enough.

“extreme performance often comes from one brutally clear”David Senra

“come up with a goal that was so simple that I could put it on a challenge coin and give it to everyone”Jonathan Ross

Principle

A founder's real job is full-time change management

Moving from engineer to founder means your primary output becomes change management.

Ross says the click for him was realizing that doing something disruptive made change management his entire job, not a side task. This reframes the founder role from technical contributor to full-time manager of human transitions.

Budget most of your energy for moving people, not building the thing.

“if I was gonna do something disruptive, my job was full Time Change Management”Jonathan Ross

Principle

The first principle of change management is to make it feel like it isnt a change

Reframe change so the goal people track stays constant.

Ross explains that whether someone experiences a shift as change depends on their altitude of focus: if they are maximizing tokens-per-second, changing the chip is not a change; if they are focused on the chip itself, it is. Change management is giving people enough context that their objective looks unchanged.

Give people the higher-altitude goal so the tactical shift feels like continuity.

“the first principle of change management is to make it feel like it isnt a change”Jonathan Ross

“People do not like change. No human being likes change”Jonathan Ross

Principle

Winners dont get more luck; they seize the luck they get (Return on Luck)

Advantage is the conversion of luck, not the quantity of it.

Ross took the Jim Collins concept and observed it directly: he missed the GitHub/LLM code-completion opportunity twice by letting his team talk him out of it, then seized it the third time by doing the arithmetic himself. He pairs Return on Luck with being positioned to see the future first.

When you feel in your bones an opportunity is right, do not let the team talk you out of it.

“the most successful companies dont have more lucky events. They just seize on that luck better than other companies”Jonathan Ross

“I had multiple lucky opportunities that I didnt seize”Jonathan Ross

Frameworks

Reusable systems and operating models — including when they help and when they break.

Framework

People Spec: a versioned written spec for the humans you hire

Write and version your hiring criteria like a product spec.

Groqs people spec (called data rock) had version numbers and was framed in positives such as Return on Luck and poetic design (semantic density: say much in few words), each carrying an implicit negative version. Ross treats undocumented criteria as a guarantee you will not hire what you claim to want.

Ship a versioned people spec; iterate it like product requirements.

“very much like you have a product spec, we had a people spec, it had version numbers”Jonathan Ross

“If you dont write down what youre looking for in people, youre not gonna hire that”Jonathan Ross

Framework

Grow-vs-Select: hire by looking for negatives, develop by showing positives

Interview to find disqualifying negatives, not confirming positives.

Ross uses a versioned people spec framed in positives (Return on Luck, poetic design) each with a negative counterpart (squander luck, maximalist design). He realized he had been hiring wrong after watching a head of HR who was excellent at spotting people problems and removing them. Grow with positives; select against negatives.

Separate the grow mode (show the path) from the hire mode (hunt the flaw) into distinct mental states.

“The biggest flip in my hiring was when I went from looking For Positives, which is what you do when youre trying to grow talent to looking for negatives, which is what you do when youre trying to select talent”Jonathan Ross

“what youre really hiring for is to avoid those negatives”Jonathan Ross

Framework

Keynesian beauty contest explains VC herding and coastal differences

VC herding is rational under a consensus-payoff game, so raise where the game favors you.

Ross explains that Groq struggled on the West coast because early VCs fell out of favor and the lemming dynamic meant one pass cascaded into universal passing; they closed with East-coast crossover funds that ignore what other VCs do. He adds the game has broken: startups now have enough capital, so extra money is no longer the advantage it was in the beauty-contest model.

If West-coast lemmings pass, do not conclude you are wrong; go to investors who run their own analysis.

“typical West Coast VCs are more like lemmings and typical East Coast VCs all think that theyre smarter than each other”Jonathan Ross

“the determiner of the the most beautiful model is not whos most beautiful, its who has the most money put on them”Jonathan Ross

Framework

GPU+LPU complementarity: defeat bottlenecks by combining hardware, not splitting stages

Combine complementary tools to defeat bottlenecks that no single tool can.

Ross uses the 18-wheeler-plus-van logistics analogy: you want both. Most people wrongly split pre-fill (reading) onto one chip and generation (the hard thinking) onto another; Groq instead assigns attention to GPU and weight-application to LPU, defeating bottlenecks across the whole decoder layer. This complementarity is what triggered the Nvidia partnership.

Map where each bottleneck lives and assign the tool that beats it, rather than forcing one architecture.

“if you were building out a logistics network for the United States, and I told you you could have either 18 wheelers or you know, vans for last mile delivery, which one would you pick? And the answer is both”Jonathan Ross

“you put these two things together and you defeat the bottlenecks across all of the different”Jonathan Ross

Framework

Intentional Leadership: state I intend to rather than ask should I

Declare intent to filter pessimism while preserving real objections.

From David Marquets Turn the Ship Around: the USS Santa Fe went from worst to first in readiness by shifting from command-and-control to intentional leadership. Ross adopted it because it fit his autonomous style; his three missed Return-on-Luck opportunities came from asking for opinions, which invited pessimism. On the third he simply stated intent and the team jumped in with how.

Replace should we with I intend to and watch pessimism drop while genuine objections survive.

“if you express Intentional Leadership, you say, I intend to do this, people dont tend to offer their opinion. But if its very wrong and theres a reason they will push back”Jonathan Ross

“I Intend To move the boat down to 500 feet, then all of a sudden someone would say, wait, the hatch is open”Jonathan Ross

Signals

What appears to be shifting, for whom it matters, and what happens if you ignore it.

Signal

Code is becoming almost free, unleashing a wave of individual founders

Zero-marginal-cost code shifts the founder constraint from capital and talent to taste.

Ross compares this to the shift from scribes to mass literacy: as code approaches free, professional engineering shifts to implement-experience-reiterate, and people with good taste who never could code (like his EA building live apps) can now create valuable software. He predicts an enormous number of individual founders.

Position for a surge of taste-driven solo founders as coding ceases to be a barrier.

“code is becoming almost free. Its, it, the marginal cost is approaching zero”Jonathan Ross

“youre gonna see individual founders without large teams creating very valuable companies that solve real problems for people”Jonathan Ross

Signal

Agentic micropayments will make the number of payments skyrocket

Agent-to-agent commerce will explode payment volume once micropayment rails exist.

Ross notes payments are not built for agents yet: his hobby project needed phone numbers from Twilio and he had to prove he was human, a friction an agent with a budget would bypass. He frames micropayment enablement as the unlock for a skyrocketing number of payments.

Build or position for agent-native payment infrastructure; volume will decouple from human action.

“if You Can make micro payments, the number of payments is gonna skyrocket”Jonathan Ross

“payments arent really built for this yet”Jonathan Ross

Signal

A year of massive up-leveling is coming for anyone who wants to learn

On-demand question-driven learning will rapidly up-level anyone motivated.

Ross argues the AI age is about asking questions, and the ability to ask in the moment of wanting to learn removes the force-feeding problem of traditional education. His advice for kids: stop teaching them to answer questions, start teaching them to ask, and revamp curricula around real community problems.

Build learning (and curricula) around asking questions about real problems, not answering pre-set ones.

“looking forward to a year of massive up-leveling for anyone who wants it”Jonathan Ross

“The ability to ask questions in the moment when you wanna learn something is gonna fundamentally change education”Jonathan Ross

Opportunities

Only included where there is a buyer, a real wedge, and a plausible revenue path — not vague idea theater.

Opportunity

Software creation for taste-rich non-coders

Enabling taste-rich non-coders to build software opens a large new founder market.

Ross EA already builds live, source-pulling apps for his trips, impossible for a non-coder before. He predicts many people with good taste who know what good is (but never could code) will now create valuable software, implying opportunity in the tooling that serves them.

Target the taste-rich, non-technical builder: they are the newly-unlocked creator segment.

“My ea creates software applications now, like when I go on a trip, she creates a little app which I can click through and it tells me what the weathers gonna be”Jonathan Ross

“a lot of people are going to get access to being able to create software to solve problems who wouldve never had the technical capabilities before, but who wouldve had good taste”Jonathan Ross

Opportunity

Agent-native payment infrastructure for budgeted autonomous spending

Payments rails built for budgeted agent spending are an unfilled, fast-growing market.

Ross first-hand friction (proving he was human to Twilio for phone numbers) illustrates the gap: an agent with an allocated budget should transact within it without human proof. As agent-to-agent commerce explodes, whoever provides the budget-scoped payment layer captures a market that grows with total agent activity.

Build the budget-and-spend rail agents need before human-proof payment friction throttles agentic commerce.

“If on the other hand I could have just done, if I had allocated a budget to the AI and it could have just, you know, used that budget, it wouldve spent it and I wouldnt never known”Jonathan Ross

Lessons still worth keeping

Useful takeaways that did not fully clear the bar for durable principle status.

Lesson

Shadowing a 2000-person leader gave Ross the confidence that transformed his leadership

Shadowing calibrates and unlocks a founders latent confidence.

At ~35 people, Ross shadowed a leader running 2000; in every meeting he silently decided what he would do, and each time it matched the experienced leaders call. Realizing he had been making the same decisions, he began acting with confidence and people followed his direction far more, without any change to the decisions themselves.

Engineer a shadowing opportunity to calibrate and unlock your own decision confidence.

“I got to shadow someone who was running an organization of 2000 people”Jonathan Ross

“at the end of it, each time when he would make his decision, its exactly what I wouldve done”Jonathan Ross

Lesson

Ross was a terrible leader for 3-4 years by delegating autonomy to people who could not handle it

Match who you hire to how you naturally lead, or execution stalls.

Ross says learning to manage cost Groq three to four years. His instinct was to delegate and set high-level direction, but he entrusted autonomy to people used to being told what to do; things stalled, and his out-of-character commands were rejected. The fix was hiring autonomous people who fit his delegating style.

Hire to your leadership style; a delegator must recruit self-directed people or nothing moves.

“I was a terrible leader. I was one of the worlds worst leader when I started”Jonathan Ross

“I didnt hire people who could operate autonomously, but I was naturally someone who would delegate and give autonomy”Jonathan Ross

Lesson

AlphaGo jumped ELO and found Move 37 only after switching from GPU to TPU

Hardware speed produced a measurable jump in model intelligence.

DeepMind, fearing a loss, moved a Go model to Ross TPU chip 30 days before a world-champion match; ELO jumped dramatically (roughly 3,200 to 3,900) on the same model, and Move 37 (a one-in-10,000 move) surfaced because the deeper search became reachable. Replayed on GPUs of the era, the model never found it.

Use compute speed as a direct lever on model capability, evidenced by the AlphaGo ELO jump.

“AlphaGo running on GPUs had an ELO score of about 3,200”Jonathan Ross

“when we went back and played it on GPUs, it never found that move because it was too deep in the chain”Jonathan Ross

Lesson

Putting fast inference on the internet triggered viral adoption after years of blank stares

Let people feel it; demonstrations to others do not convert, personal use does.

For years Ross hit brick walls explaining fast inference (people asked why they would need an LLM faster than they can read). Recalling that an Anthropic demo three months before ChatGPT drew no reaction because the audience only watched, he concluded the only path was to put it online. Usage skyrocketed after a viral X video; his own demo in Norway even felt slow because servers were saturated.

When a benefit is only believable when felt, ship the felt experience rather than the argument.

“Someone had posted on, on XA video of an LLM running on GR that was just running super fast and it was viral”Jonathan Ross

“there was no possible way, no matter what we showed people, for them to accept that fast inference was gonna be helpful unless we let them try it”Jonathan Ross

The Plays

Try these this week

Verb-first executable actions — each one tied to a stated outcome in the episode.

Maintain a versioned people spec framed as positives with negative counterparts

Outcome: Codify each hiring trait with its negative twin so you can select against the negatives.

Context: Groqs people spec (data rock) was versioned like a product spec and framed in positives: Return on Luck, poetic design (semantic density, every word matters). Each has a negative version (squander luck, maximalist design). The play enables the grow-vs-select flip: hire by hunting the negative counterpart.

“we framed the people spec in Positives. Things that you look for, like Return On Luck”
— Jonathan Ross

maintained continuously, versioned per

Before you start

· clarity on the traits that matter
· discipline to maintain versions

Shadow a far more senior leader to calibrate and unlock confidence

Outcome: Shadow a senior leader and silently predict their calls to calibrate your judgment.

Context: At ~35 people, Ross shadowed a leader running 2000, attending every meeting silently and deciding what he would do before the leader decided. Each match accumulated evidence that his judgment was sound; recognizing this, he began acting with confidence and people followed, with no change to the decisions.

“I got to shadow someone who was running an organization of 2000 people”
— Jonathan Ross

weeks of sustained shadowing per

Before you start

· access to a senior operator
· willingness to stay silent and observe

Kill one-on-ones and copy everyone on every email to starve politics

Outcome: Broadcast information and copy everyone so politics has no dark corners to grow in.

Context: Ross learned this at Groq and saw it in the extreme at Nvidia, where Jensen never has one-on-ones telling people different things. When he told one person something then another, they compared notes and reached divergent conclusions; a room full of people heard the same thing. If someone flags anothers failing, copy that person so they can jump in.

“stop having one-on-ones. Have big meetings with everyone who you wanna tell something and tell them all at once”
— Jonathan Ross

ongoing policy per

Before you start

· leader tolerance for open conflict
· culture that can handle candor

Convert salary into equity to cut burn without cutting the team (Grok Bonds)

Outcome: Trade salary for equity to preserve critical talent through a cash crisis.

Context: Three weeks from running out of money, Groqs leadership drafted a layoff list; Ross saw the cuts would kill the not-yet-working product (a novel kernel-free compiler needed critical mass). Instead they ran an all-hands with WWII war-bond imagery, branded it Grok Bonds, and offered salary-for-equity. 80% participated, about half dropping to statutory-minimum salary; attrition was under 10% (possibly ~5%, better than before) and they saved roughly two months of runway.

“it was an exchange of salary for equity. And we expected that we were gonna have pretty high attrition and we actually didnt. 80% of the employees participated”
— Jonathan Ross

decided in weeks under an existential clock per

Before you start

· equity worth believing in
· transparent communication of the crisis
· legal statutory-minimum salary floor

Stamp the whole company on one metric via a challenge coin

Outcome: Distill the mission into one number and give everyone a coin with it.

Context: Ross gave every Groq employee a challenge coin reading 25 million tokens per second with a graph trending up. Chip, software, power, supply-chain and datacenter teams all connected their work to that number. The play only works if the metric is the true dominant game and crisply distilled.

“everyone at Grok had a challenge coin that said 25 million tokens per second and had a little graph of it going up and everyone knew that was the thing to do”
— Jonathan Ross

set once, reinforced continuously per

Before you start

· a genuine dominant-game metric identified
· ability to crisply distill the objective

Put everyone's hands on the steering wheel to reduce fear during a crisis

Outcome: Give people an active role in the rescue and their fear drops.

Context: Ross uses the driver-versus-passenger metaphor to explain why Grok Bonds cut attrition below 10% (possibly ~5%, better than before the crisis): letting employees participate in saving the runway gave them control, so despite real pain they felt more secure than if the fate were done to them.

“When people are passengers in a car, theyre more nervous about a windy road or a scary road, but when theyre the driver, they feel more in control”
— Jonathan Ross

the duration of the crisis per

Before you start

· a genuine role people can play
· transparency about the situation

Hire for loss bias: select people who book the win early

Outcome: Recruit for the instinct to treat an unrealized possible win as a loss.

Context: Ross noticed in architecture meetings that when someone said a change would make the chip twice as fast, the room was flat: they heard put it in the next chip. He heard: if we dont, this chip is half as fast as it could be. He hires for this book-the-win-early trait and applies the lens to product decisions, not just people.

“Theres a personality trait in people where I call it sort of booking the win early”
— Jonathan Ross

ongoing per

Before you start

· awareness of loss bias
· a people spec to encode the trait

Manufacture pressure by publicly staking your reputation on the outcome

Outcome: Stake your reputation publicly to force yourself to a higher level.

Context: Ross reads Michael Jordan (from the Founders episode) as intentionally taunting opponents so a loss would be humiliating, forcing superhuman performance. Most people keep their sights low to avoid the pain of a public miss; entrepreneurs instead declare success publicly, raise money, and put their reputation on the line to force themselves to deliver.

“hes very intentionally throwing his keys over the fence, so he has to go fetch them”
— Jonathan Ross

per commitment cycle per

Before you start

· tolerance for public risk
· an outcome worth the exposure

Phrase decisions as I intend to instead of asking permission

Outcome: Declare intent to convert reflexive objection into execution help.

Context: On the third Return-on-Luck opportunity, instead of asking whether it could be done, Ross presented a target speed-per-chip as a decision. Rather than saying it was impossible, the team jumped in with how to do it, and they hit the numbers. He credits David Marquets Turn the Ship Around and the USS Santa Fe turnaround.

“I literally said, I, I, I put a presentation together, I said, were gonna get to this particular speed per chip, and rather than people going, we cant do this, they all sort of jumped in and said, this is how we do it”
— Jonathan Ross

immediate, per-meeting per

Before you start

· a formed point of view
· psychological safety for real objections

Ship the experience publicly so people feel the benefit instead of hearing about it

Outcome: Manufacture the magical first-person moment by shipping a public try-it product.

Context: Ross recalled an Anthropic demo three months before ChatGPT that drew no reaction because the audience only watched, versus ChatGPT where everyone got their own answer. Concluding that only personal experience converts, Groq put fast inference online; a viral X video caused usage to skyrocket (his own Norway demo even felt slow from server load).

“the only way were gonna get people to understand the value of speed is if we just implement this, put it on the internet. So we did”
— Jonathan Ross

launch then ride the viral curve per

Before you start

· a working product
· capacity to absorb a usage spike

Always prototype on a personal hobby project before bringing it to work

Outcome: De-risk new capabilities by learning them first on a personal hobby project.

Context: Ross builds cutting-edge things on his personal computer and spun-up GCP/AWS servers (travel-route apps, a personalized daily brief, math apps) precisely because it avoids risk to the work code base. He uses AI extensively at work but always prototypes on a hobby project first, building intuition before organizational application.

“I Always Start With A Hobby Project before I bring it to work”
— Jonathan Ross

ongoing personal practice per

Before you start

· personal infrastructure separate from work
· time for side projects

Decision Moments

Actual decisions, real outcomes

Specific decisions narrated in the episode with their outcomes and transferable lessons.

Groq was about three weeks from running out of money, pre-product-working (still needed a novel kernel-free compiler to hit critical mass). The leadership team had drafted a layoff list, but the people slated to be cut were exactly the talent required to make the product work — so layoffs would have killed the company.

Did: Rejected the layoff plan after doing the burn math, and instead launched Grok Bonds: a voluntary salary-for-equity exchange, framed with WWII war-bond imagery at an all-hands, letting employees cut salary down to statutory minimum to preserve the team and reduce burn.Outcome: 80% of employees participated, about half dropping to statutory-minimum salary; attrition stayed under 10% (possibly ~5%, better than before the crisis). Saved roughly two months of runway; they had three weeks of cash when they raised. Without it, the company would have gone out of business.

In a cash crisis, protect the critical talent burn-math cannot afford to lose: trade salary for equity and let the team participate in the rescue (hands on the steering wheel), which lowers rather than raises attrition.

Part of an emerging decision pattern across multiple episodes

Groq had spent 3-4 months integrating GPU and LPU hardware to defeat heterogeneous LLM bottlenecks. Ross initially did not think it was a big enough deal to propose to Nvidia. His COO Sonny autonomously had the idea to put the chips together; the team then figured out the assignment (attention on GPU, weights on LPU). They went to Jensen to buy ~100k GPUs to deploy themselves.

Did: Let the autonomous team run with Sonny's combine-the-chips idea, implemented it, and — rather than hiding it — showed the working GPU+LPU system to Nvidia because they wanted to become a GPU customer. Jensen saw it and proposed making it available to all Nvidia customers.Outcome: A ~$20B partnership closed in a three-week window from the presentation to money wired — the biggest deal Nvidia had ever done by nearly 3x. West-coast VCs had passed on Groq and missed it. The idea originated not from the founder but from an autonomously-operating COO and team.

Autonomy plus willingness to show your work to a potential partner (not fear of exposure) can surface founder-scale opportunities the founder himself underrated; move fast because technology has real opportunity cost to waiting.

Part of an emerging decision pattern across multiple episodes

When LLMs first emerged, the GitHub CEO called Ross wanting LPUs for code completion because they could not get GPUs. Ross believed in his bones it was ideal for the LPU, but his team said it could not run on the chips, focusing on missing GPU features that did not actually matter for LLMs. This recurred a second time on another LLM deployment opportunity.

Did: Twice deferred to the team and let them talk him out of the opportunity against his own conviction (squandering luck). The third time, he refused: he did the arithmetic himself, presented a target speed-per-chip as declared intent (I intend to), and the team jumped in with how — hitting exactly the performance numbers.Outcome: Missed being the original LLM host for Microsoft/OpenAI (a very different outcome), but on the third opportunity, seizing it with intentional-leadership phrasing produced the exact performance targets and kept Groq ahead of the fast-inference curve.

Return on Luck is about conversion, not luck volume: when you feel an opportunity in your bones, do not let a team anchored on the wrong reasons talk you out; declare intent rather than ask for opinions.

Part of an emerging decision pattern across multiple episodes

Tensions surfaced

Contradictions and trade-offs the episode raises — judgment calls a thoughtful operator has to navigate.

Tension

Under-constraining versus over-constraining the goal

Constrain the what crisply, leave the how open.

Ross frames both errors as failures of distillation. If you cannot crisply state what you are accomplishing, you will either over-constrain (no room to innovate) or under-constrain (people flounder). The 25M-tokens-per-second coin is his resolution: maximally clear objective, maximally open method.

Invest in distilling the one clear goal; it is what lets you under-specify the method safely.

“If you arent able to very crisply distill what youre trying to accomplish, then either youre gonna over constrain or under constrain the people”Jonathan Ross

“you must not over constrain the goal”Jonathan Ross

Tension

Too much confidence versus too little confidence

Think like the underconfident, act like the confident.

Ross frames it as two failure modes: too much confidence (acts without thinking) and too little (thinks but cannot act). The low-confidence person has an analytical edge but must still learn to act decisively so teams get the conviction signal. Resolution: keep the deep analysis, but arrive at decisive action.

If you are analytical, keep it, but train yourself to convert analysis into confident action.

“Many people have too much confidence, some people have too little confidence”Jonathan Ross

“if you have too little confidence, youre probably the kind of person who thinks through things a lot more. But you need to still get to a point where you act with confidence”Jonathan Ross

Tension

Real feedback versus unnecessary pushback

Get true objections without inviting reflexive pushback by declaring intent.

Ross names the arc: early leaders drown in pushback, senior leaders starve for feedback. The lever is subtle phrasing. Asking should I invites pessimism; stating I intend to removes the invitation for opinion but still prompts people to raise a real, specific problem (the open hatch). This balances the two failure modes.

Tune the phrasing of decisions to dial feedback between too much pushback and too little.

“early in their career. They get way too much pushback later in their career. They dont get enough feedback”Jonathan Ross

“I was inviting pessimism by asking for peoples opinion”Jonathan Ross

Corpus connection

Where this episode fits for retrieval

What kinds of decisions this briefing is best pulled into.

Primary decisions

• hire
• strategic-bet
• organizational-design
• fundraise