[Jun 29–Jul 5] | Electricity Beats Model Quality · Token P&Ls · Taste Moats

Electricity is becoming the AI moat nobody budgeted for

The strongest thread this week was not model quality. It was physical throughput. Bloom Energy framed AI as an electricity business, Etched framed inference as a hardware specialisation problem, and Groq framed speed as an architecture problem where power, memory, chips and deployment all sit inside the product experience. Once AI moves from demos to mission-critical work, uptime and latency stop being facilities concerns. They become customer promises.

That makes the usual “which model wins?” debate feel too narrow. KR Sridhar’s claim that electricity, not AI models, will decide the winners is deliberately provocative, but the evidence is no longer theoretical: Bloom says it has around $20bn of backlog, roughly 1 GW of manufacturing capacity today, more than 2 GW expected by year-end, and delivered 50+ MW for Oracle in 55 days. In that world, the bottleneck is not just intelligence. It is permitting, grid access, cooling, redundancy and time-to-power.

The contrarian read is that the deeper AI moat may sit outside the model layer. If every serious company can rent good-enough intelligence, the advantage shifts to whoever can deliver reliable, low-latency compute where and when customers need it. Distributed power, transformer-specific ASICs and latency-first inference stacks are not back-office infrastructure. They are how AI products become commercially real.

Operators should run their AI roadmap through a power lens. Where does compute live? Which workflows break during outages? Which products depend on sub-second latency? Which geographies are slowed by grid or permitting constraints? For infrastructure-heavy businesses, time-to-power now deserves the same status as time-to-build.

Token spend needs a P&L, not a hype budget

Several episodes converged on the same uncomfortable point: AI revenue and AI usage no longer prove good economics. Benchmark described the “death of spreadsheet investing” because classic SaaS heuristics are breaking. Sierra’s Clay Bavor said strong engineers at top AI-native companies can spend more than $100k a year on tokens, with tokens potentially becoming around 20% of developer compensation economics. Coinbase reportedly cut AI spend by 50% while usage rose.

The old software model rewarded high gross margins, clean seats and low services load. AI bends that. Benchmark’s P x Q x M frame is useful because P can rise sharply when software sells completed work rather than seats, but M can fall when inference is expensive and implementation is heavy. A company can look exciting on revenue and still be fragile if every new customer drags a large token bill behind it.

The sharper read is that high gross margin may sometimes mean underuse, not quality. If AI is genuinely doing work, it may consume meaningful compute. The question is not whether the bill is large, but whether the work completed justifies the cost of intelligence. “Adoption” is a vanity metric unless it is tied to velocity, margin, revenue, deflection, throughput or fewer human hours in a real workflow.

CFOs and product leaders need a token P&L by use case. Separate frontier-worthy tasks from cheap-enough tasks. Route low-stakes work to open models or fine-tunes, reserve premium models for high-value reasoning, and price against outcomes rather than seats where the product replaces labour. The next board question will not be “are we using AI?” It will be “which AI spend compounds?”

Speed is now a product boundary

Etched and Groq both made latency feel strategic rather than technical. Etched breaks the problem into first-token delay and per-token speed; Groq talks about compute-bound and memory-bound workloads, and built an operating story around a 25 million tokens-per-second goal. Sierra adds the application-level version: faster deployment, tighter feedback loops and AI-assisted engineers who can be 3x to 20x more productive.

The key point is that agentic systems multiply latency. A slow chatbot is annoying; a slow agent is structurally limited because one model call triggers another, which triggers a tool call, which triggers more reasoning. Every second saved expands what the system can attempt. Faster inference can make the same model feel smarter because it enables more search, more retries, more context gathering and more downstream work before the user loses patience.

That implies many “AI product” ideas are really infrastructure bets in disguise. Voice, robotics, live search, support agents, coding agents and workflow automation only become categories when the interaction loop feels instant enough. If a use case is not tolerably fast, it may not be commercially real yet, even if the model looks clever in a benchmark.

Product teams should measure latency as a conversion and completion variable, not a dashboard vanity metric. Track first-token latency, per-token speed, task completion rate, abandonment and the number of downstream actions an agent can finish in a session. Speed can change category economics, not just UX polish.

Open source is becoming procurement leverage

The open-versus-frontier debate looked less ideological this week and more like procurement strategy. Coinbase cutting AI spend while usage rose suggests open models are now budget-control tools for normal software companies, not hobbyist substitutes. Sierra’s stack makes the same point from the builder side: it uses open-weights models plus proprietary fine-tunes rather than spending billions to pre-train a foundation model.

The market is likely to split by task, not tribe. Frontier models still matter where better reasoning is deeply monetisable: coding, science, legal work, materials, high-stakes customer interactions. But if open models can handle 80% of repetitive work, frontier models become premium options rather than defaults. That weakens pricing power at the model layer and increases the value of routing, orchestration and evaluation.

Anthropic’s complaints about Chinese distillation are interesting for the same reason. The public framing is safety, IP and national security; the commercial subtext is frontier economics. If policy turns Chinese models into a procurement risk, US open-source alternatives may benefit. If it does not, open models keep compressing the price umbrella under frontier labs.

Operators should build model optionality before they need it. That means clean abstractions, model routing, vendor switching, evaluation harnesses and explicit rules for which workflows deserve premium intelligence. Lock-in may feel convenient while budgets are experimental. It will feel expensive once token spend becomes a board-level cost line.

Forward deployment is the new enterprise AI distribution

Sierra, Benchmark and Groq all pointed towards a less polished but more durable enterprise AI truth: the product is not separable from deployment. Sierra’s early design partners included Olukai, SiriusXM, Sonos and Weight Watchers, and it later took large customers such as Next and Cigna live quickly. Benchmark called the broader pattern “FDE is the new PLG” and “the Palantirification of everything.”

That sounds ugly if you grew up on clean SaaS metrics. Services-heavy onboarding used to be a red flag because it implied low margins and bespoke work. In enterprise AI, the same work can be the learning loop. The vendor that embeds inside customer workflows learns the edge cases, builds the context layer, tunes the agent and turns implementation pain into reusable product capability.

The contrarian read is that the messy parts may be the moat. Enterprise buyers do not just want an interface to a model; they want outcomes inside their existing systems, permissions, policies, customer data and operating rituals. Owning that context matters more than owning a narrow feature. The first sale is often a research programme disguised as onboarding.

If you sell enterprise AI, deployment should report into product strategy, not sit as a post-sale afterthought. Build shared context infrastructure, shorten the loop from customer discovery to product change, and decide which implementation patterns become platform assets. PLG taught software companies to remove humans from distribution. Enterprise AI may reward the teams that put the right humans back in.

Taste is the scarce layer when output is cheap

Dylan Field’s “permanent underclass of zero taste” line is the cleanest warning for anyone using AI to make product, design or marketing cheaper. His argument is not that AI makes design irrelevant. It is that AI makes average output effortless, which increases the premium on judgment, point of view and the ability to steer work away from the mean.

This matters because many teams are collapsing PM, design and engineering into one fast solo workflow and mistaking speed for taste. FigmaWeb was the useful counterexample: not just prompting, but a workflow for constraining, composing and improving model output. Years after GPT-3, we are still mostly typing into boxes. That suggests the interface revolution is unfinished.

The provocative read is that “execution is cheap” is only half true. Shipping will get easier; deciding what should exist, what should be excluded, and how a product should feel will get harder. Brands with no point of view will become easier to imitate. Products with no interaction philosophy will drift towards the model average.

Operators should audit for average-ness. If an LLM can produce your homepage, onboarding, sales deck or launch copy with little loss, the problem is not the model. It is the absence of a view. Keep humans explicitly in the loop where empathy, strategy, taste and trust are the work, and design workflows that help people steer outputs rather than merely request them.

AI leverage makes vague management more expensive

The management thread across Groq, Sierra and Figma was simple: as leverage rises, ambiguity gets costlier. Jonathan Ross described Groq’s challenge coin around one metric, 25 million tokens per second. Sierra runs a six-week board cadence, memo-first materials and fast reprioritisation. Dylan Field kept returning to empathy, judgment and craft as human layers that cannot be delegated casually.

The useful management insight is not “be clear”, which is too generic. It is that high-leverage organisations need one dominant game. If every team optimises its own local KPI, AI does not fix the company; it accelerates the fragmentation. Ross’s “return on luck” framing is apt: companies do not win because they get more lucky breaks, but because they recognise and act on them faster.

Hiring also gets less fuzzy. Ross moved from hiring for positives to screening for failure modes: people who squander luck, over-design, or create political drag. Sierra now evaluates candidates by giving them a prompt, $150 of token budget and a coding-agent setup. The point is not whether someone can talk about AI. It is whether they can use leverage with judgment.

Founders should make anti-traits explicit, replace vague consensus-seeking with clearer statements of intent, and align the company around one game that survives strategy shifts. In a world where small teams can do more, the limiting factor is less likely to be headcount and more likely to be unclear judgment at the top.