Recommendation · MoneyPer-task token costs have fallen so much in the last 6–8 months that it is now far harder for someone to 'token max' and blow up company costs; the real danger point was 6–8 months ago. Let team members 'go nuts' with expensive tokens early — doing the weird, expensive stuff first is how you discover the lessons you can scale back with later. Expect frontier-model costs to keep falling as competition from open source and closed source drives costs down.
Latent.Space · AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman & CJ Hess
Recommendation · MoneyTime new product bets to model capability jumps — their agent product hit product-market fit quickly because a simple insight coincided with models (Opus 4.5 era) suddenly becoming more agentic.
described
The a16z Show · What Happens When Al Agents Can Spend Your Money?
Report · MoneyThe speakers describe Opus 5.5 as the preferred model in the office right now: it is good enough, fast, and cheap, hitting all checkboxes. Previously they used two to three models depending on whether they were cheap and fast, really smart, or could solve specific problems; now Opus covers the full range. The team still reaches for tools like Codex when needing to do something in the computer use world, and using Opus 5.5 as the driver with Codex to review — working back and forth on the machine — still produces the best results. Costs have also been driven way down from Opus 5.5.
Latent.Space · AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman & CJ Hess
Recommendation · MoneyRandy Hunt recommends asking students or mentees to write a five- or ten-year plan — a vision-setting exercise that many people, including Debbie Millman, have found transformative. He notes such exercises were novel imports into design education from coaching or therapy, and credits Milton Glaser with innovating by bringing them into design practice.
Design Better · Randy Hunt: Notion and Etsy design leader on why the best design teams stay small
Recommendation · MoneyRandy Hunt recommends that when you feel stuck working on something, ask: have we ever talked to customers about this thing?
confident
Design Better · Randy Hunt: Notion and Etsy design leader on why the best design teams stay small
Report · MoneyCarmen Li points to a secondary server marketplace on Compute Exchange with live inventory of used and refurbished GPUs, including 409Ls, B200, and AMD hardware.
AI Engineer · Buying GPUs Should Work Like Buying Stocks — Carmen Li, Compute Exchange
Creator perspective · MoneyCarmen Li explains that implied flow rates are calculated from the term structure using methodology similar to Treasury futures curves, because there is no GPU futures trading. The market consensus view from implied rates is that GPU prices could go up over the next 36 months. However, she cautions that market consensus does not mean it is right.
Methodology description and a hedged forward-looking expectation, explicitly not a certainty.
AI Engineer · Buying GPUs Should Work Like Buying Stocks — Carmen Li, Compute Exchange
Creator perspective · MoneyThe pitch for vertical AI is focus — they care about their last mile in terms of the software the way you care about the last mile for your customers. The answer to why vertical AI companies win so far is access to data that ChatGPT will not have, which is a moat; Anthropic and OpenAI are also important acquirers of the same data sources but are not as dedicated as vertical AI companies will be. Harvey has Tabular Review and Legora also has it — customized affordances and UIs because they only serve their vertical. Harvey for lawyers and Open Evidence for doctors are the classic examples of vertical AI companies winning their domains.
the host's interpretation
Latent.Space · AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman & CJ Hess
Creator perspective · MoneyExpect a small set of companies to own the agent interface, with everyone else plugging into it via APIs/MCPs — you cannot stack unlimited $20/month agent subscriptions. Salesforce is embracing this 'interface owner' path: providing APIs and MCPs so customers plug external agents into Salesforce rather than using Salesforce's own agents.
Latent.Space · AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman & CJ Hess
Report · MoneyDrawing on roughly 100 enterprises Tenex talks to, the speakers report that 90% have rolled out either Claude or Codex to their employees. Roughly 60% have reached the training and enablement milestone — which never gets the love it should but is arguably as important or more important than the technology itself. The number that have rolled out coding agents to the engineering product org is closer to 50%, not 70–80% as most people assume; every other enterprise either hasn't rolled out a coding agent or what they use is closer to not having one. Agents in production for non-engineering work drops to about 5%, and most enterprises with more than 1,000 employees do not have an agent in production doing work that is not building software.
approximate percentages from Tenex's own conversations
Latent.Space · AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman & CJ Hess
Report · MoneyTenex pays people by story points, with agreement on what the story points are. Some Tenex engineers can literally make millions of dollars a year while others may not make as much because they're not as productive. There is a very wide dispersion in earning capacity between the highest and lowest paid engineer at Tenex — roughly a 4x difference.
speaker says 'probably' and 'something around' for the 4x figure
Latent.Space · AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman & CJ Hess
Creator perspective · MoneyBe generally bearish on vertical-specific AI solutions (e.g., Harvey for lawyers, Open Evidence for doctors) because enterprises have enormous nuance, inertia, and systems to wrangle, and buyers now expect one-of-one software. Agent-as-product vertical AI tools will struggle unless they have a strategic moat (compliance, data) — because you can spin up your own quickly, models keep compounding, and tasks fall into general models. Vertical AI can survive via a data moat (access to proprietary data general models won't have) or through focus: customized affordances, UIs, and last-mile care for a single profession. Vertical AI products that change the process and put an agent on top of an app/tool built around it can still be successful.
Latent.Space · AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman & CJ Hess
Creator perspective · MoneyThe next time an agent makes flyers, they should automatically look like the new design because it should have noticed the design change for you — you shouldn't even have to ask. And when agents can do video effectively, they should proactively suggest making video assets so you can advertise on Instagram.
hypothetical product expectations presented by the speaker
The a16z Show · What Happens When Al Agents Can Spend Your Money?
Creator perspective · MoneyWriting prep notes for every meeting is a long-tail task that becomes worth automating because an LLM can do it for about a penny. An executive assistant is not going to spend time writing perfect prep notes for every 30-minute meeting — it's just not a good use of time.
cost figure is approximate
The a16z Show · What Happens When Al Agents Can Spend Your Money?
Report · MoneyIn the example product scenario, a user asks in Town whether anyone knows someone on the finance team of a potential customer, and Town queries the agents of other employees to find a connection. The human-speed version is posting in Slack or Microsoft Teams, where everyone reads the message, thinks about it, and eventually answers. With agent mediation, the user gets the answer immediately because the company has decided to trust AI to mediate the security and privacy of that request within the company.
hypothetical product scenario presented by the speaker
The a16z Show · What Happens When Al Agents Can Spend Your Money?
Creator perspective · MoneyThe speaker thinks of their product as a token allocation machine that gives humans options about where to apply tokens and lets the human choose the outcome. As tokens get cheaper, there are more and more places where applying tokens becomes worthwhile that would never have been done with today's expensive tokens or with humans. Writing prep notes for every meeting is a long-tail task that becomes worthwhile because an LLM can do it for about a penny, whereas an EA would never spend time doing it for every 30-minute meeting. Meeting prep notes can include research on who else is in the meeting and whether they have said anything relevant in the last week.
the penny cost is the speaker's rough estimate
The a16z Show · What Happens When Al Agents Can Spend Your Money?
Report · MoneyWhat makes someone stick with Town: they've connected their email, their calendar, and ideally one really important system of record for their business, plus one custom integration. The goal is to eliminate the custom integration — the speaker is trying to get rid of it — so that as soon as you connect your system of record, the product delivers value that feels very tailored to you.
speaker's description of their own product
The a16z Show · What Happens When Al Agents Can Spend Your Money?
Report · MoneyThe speaker's company bets on starting with PLG, SMB, mid-market, and tech-forward enterprises because their automation approach scales well there, leaving handcrafted enterprise software to others.
speaker's description of their own company's strategy
The a16z Show · What Happens When Al Agents Can Spend Your Money?
Report · MoneyAmerican restaurant owners receive wholesaler price lists by email roughly once a month or once a quarter, as Excel spreadsheets or PDFs covering thousands to tens of thousands of items, with prices constantly changing.
described with 0.8 confidence; speaker's characterization of the market
The a16z Show · What Happens When Al Agents Can Spend Your Money?
Creator perspective · MoneyRandy Hunt identifies three reasons experienced designers leave talent-dense teams: reaching a natural next career chapter after accomplishing a lot, losing conviction in the vision when the thing they bought into changes, and the itch to build from zero again. Of the three, only losing vision conviction is one the team has control over.
Design Better · Randy Hunt: Notion and Etsy design leader on why the best design teams stay small
