Carmen Li recommends verifying provider claims before purchasing by checking UIDs and inventories and running performance benchmarks such as NVIDIA DGX, InferenceX, and MLperf. For forward capacity, she says providers' purchase orders and permits should be reviewed before committing to future delivery.
Have agents log their decisions and what they're doing in an opinionated, standardized way, all checked into Git — so there's a durable record of why decisions were made and the context around conversations. Don't rely on code alone as the memory of a codebase; you need docs to know why you made a decision. Keep all context in the repo, never in someone's head — this gives agents and tools superpowers by having all the context in one place, enabling more agent-driven work.
A business should own its AI context assets so thoroughly that if a better solution than the current one existed, switching to it would take only a day. Businesses want both neutral intelligence and an app-layer solution they feel comfortable they could move away from — owning the context layer with someone who makes it available to them. Companies deploying FDEs to build software for you are actually trying to make it more difficult for you to switch to someone else at any point.
speaker's recommendation and characterization of FDE vendors; the one-day figure is illustrative (possible numeric error flag)
CJ Hess phrases it to his team: you can't automate a process that does not exist. When 40 people all roughly do the same function but each does it a little differently, it is way harder to build an agent that proactively does that work in the background. Without some form of standardization, it's impossible for an agent to know whether it's doing the work right, to build evals around the system, and to deliver the value an agent doing that function would have.
Standardize a set of skills into structured documentation within a repo plus a CLI that models use to drive work — creating plans, visualizations, client readouts — so the pieces around writing code happen automatically and standard across the team. Maintain a big canonical plan document with consistent structure, but reduce enforcement of that structure as models get better.
A restaurant owner's part-time finance person should review how wholesaler prices have changed at least quarterly — and the speaker endorses running the analysis at least monthly, whenever prices change, or even daily if a wholesaler changes prices daily — to catch overpaying (e.g., a price up 20-30%) or to shift orders to different wholesalers.
advice attributed to the speaker; the 20-30% price-increase example carries a possible numeric error flag
Businesses should choose neutral AI applications that use whatever model supplier is best for them in the moment, rather than locking into one lab's product; smart businesses should think carefully about the neutrality of who provides their intelligence and who does the routing. Model supplier choice is not just about general intelligence — suppliers differ in which modality (voice, video, slides) they are best at at any given time. Choosing a neutral intelligence provider lets a business save money and de-risk ever having picked the wrong AI partner.
speaker's recommendation, endorsed; the modality claim is described
You cannot automate a process that does not exist. Before handing work to an agent, build a standardized process around the org function, then define the steps to hand that work off to an agent running in the background. Without standardization you also cannot build evals or know whether the agent is doing the work right.
A step-by-step method for acquiring GPU compute: sign up for free, browse listings, submit an RFQ with full requirements across multiple regions, and evaluate provider responses on the full rubric rather than price alone.
Randy Hunt recommends building design teams where designers are highly technical and ship code in the production application — a 'software design team.' He says it's hard for him to imagine building a software product any other way, and he built such a team at Etsy in 2010, long before coding agents made it trendy.
Sujee argues KVCache is probably one of the biggest ROI optimizations in doing inference, because creating tokens is expensive, especially one at a time. Once you generate a token, it doesn't change, so there's no need to keep regenerating the same token over and over again — the idea is to cache tokens so the next time you need to look up, you don't have to regenerate it, you just look up the cache. Nebius has seen speedups anywhere from 5x to 10x, not just 5 to 10 percent.
Carmen Li advises that if you have servers on your balance sheet with financing obligations for the next five to six years and reserve contracts for only the first three, you are exposed in years four, five, and six. To hedge out that exposure, enter swaps or block trading. Even for the first three years covered by reserve contracts, there is counterparty default risk if the reserve contract counterparty fails, which can also be hedged via block trading around indices; block trades have a floating rate.
Creator's advice directed at the audience; counterparty risk discussion is her characterization.
Carmen Li recommends matching contract type to need: reserve contracts for guaranteed compute with verified benchmarks, forward contracts for GPUs starting at a future date, and token forward contracts for dedicated token deployment with guaranteed SLA and fixed quantization/precision.
The next enterprise AI adoption challenge after broad LLM rollout is permissioning and RBAC: as every employee gets Claude or Codex, they need controlled access to certain data. Give every AI-using employee a data-permissioning gateway (an RBAC control panel of what each person can and cannot read) — otherwise AI access to company data becomes a nightmare for CIOs and CISOs. Tenex started by building bespoke data-permissioning gateways for companies, then productized that into their own gateway solution.
Carmen Li describes how last year the implied forward curve for GPU compute was completely downward sloping, meaning longer-term contracts were cheaper per chip per hour. Starting early this year, the whole curve shifted up and became contango, implying people assume prices will rise down the line and aren't getting enough discount for longer-term contracts. The term structure has now flattened, meaning people will either get penalized or awarded by a longer-term contract. For B200, last year the curve was contango and this year the whole curve shifted up — every single term is more expensive. In platform rates, people similarly think the price is going to go up.
Market observations and interpretations from the speaker; some atoms were flagged as possible numeric errors.
Frame enterprise AI adoption as a sequence of milestones in a race and use this ladder to diagnose where a company actually is rather than assuming uniform progress. Benchmarks from ~100 enterprises: ~90% have rolled out a general-purpose LLM (Claude or Codex) to employees; ~60% have done training and enablement; only ~50% have rolled out coding agents to the engineering org; and only ~5% have agents in production for non-engineering work. Don't assume an enterprise has coding agents just because they've rolled out LLMs — roughly every other enterprise either hasn't rolled out a coding agent or what they use is closer to not having one.
People who tried AI two years ago ossified that version of reality; if they try it again now, the results are dramatically different. The speaker has run this experiment about 10 times. A Harvard professor dismissed AI based on trying it a few years ago until the speaker had ChatGPT rewrite her 20-year-old dissertation with everything current, and she was amazed. Silicon Valley judges AI by where it could go (the slope of the curve), while most people judge it by what it is today — which is why SV sometimes oversells reality but also sees the trajectory others miss. In the last two and a half months, computer-use and browser-use assistants can do things that were near-impossible six months ago, with success rates jumping from roughly 30% to roughly 90% of the time.
speaker-reported figures; success-rate numbers flagged as possible numeric errors
A restaurant owner can prompt their way to a monthly wholesaler price-change analysis in about five minutes; it costs roughly a dollar a month to run, and even if it finds nothing in 11 out of 12 months, in the month it does find something it can uncover hundreds if not thousands of dollars in savings.
cost and savings figures are the speaker's rough estimates ('like a dollar a month', 'hundreds if not thousands'); five-minute setup and dollar cost carry possible numeric error flags
Randy Hunt argues for keeping design teams small and dense with strong people: small size reduces communication overhead, constrained bandwidth becomes a forcing function for prioritization, and a team that is too small becomes a delivery-only production engine, which he calls a fail state.
Carmen Li argues that for capital markets and resource allocation the key variable is efficiency. Buying stocks via NYSE or NASDAQ doesn't guarantee the best price, but it is more efficient and liquidity is assured. She applies the same logic to GPU procurement: when you need compute within 48 hours on reserve contracts, going to an exchange with verified compute resources is the easiest way, analogous to going to NASDAQ to buy stocks. Calling 300 salespeople is impractical when you need compute in 48 hours, though talking to cloud salespeople regularly is still valuable because they know a lot of things.
Consequential creator advice requires independent verification before Today can promote it.
No published protocol matches this selection.
The creators behind most of what cleared.
195 cleared insights match
99 cleared insights match
85 cleared insights match
Choose sources → to get a library of your own and a daily brief.
