Apoorv Agrawal

ChatGPT is at 900M weekly actives (“about 10% of the world coming to us now, 90% left to go”), and Nick Turley allocates “all my points” to long-term retention as the north star: “things like revenue, they follow from that.” The proof point: moving GPT-4-class intelligence from behind the paywall to free (4o) was “totally revenue positive and retention positive.”

Historical growth splits one-third / one-third / one-third: classic friction removal (killing the login wall — “Sam will say I told you so”), core product built jointly with research (search, personalization, post-trained into the model), and pure model gains — both step changes (3.5→4, 4 paywalled→4o free) and unsplashy iteration like 5.3 and 5.4.

The next billion requires going beyond chat: today’s product is “a raw appliance… too much like a computer terminal,” and the unlock is to “productize reasoning in a way that works on people’s behalf without them even knowing” — long-horizon tasks users never encounter as a concept. Actions plus proactivity compound into the super assistant; the codebase is literally named “SA server.”

Agent timing is the tell: ChatGPT agent was “slightly too early” — without escape velocity “users don’t learn to trust it, they don’t even try” — but general-purpose agents are near the partial-credit threshold where hill-climbing begins: “on task I think we’re close.” Codex already has escape velocity (“so many engineers who don’t open their IDE like ever”), and quantitative knowledge work follows because it’s “testable… very RL friendly.”

Pricing will evolve: power users are getting “almost too much value,” and “having unlimited plan is like having unlimited electricity plan… there’s a reason you can’t buy that.” Ads are framed as an access play for markets without credit cards, principles (answer independence, privacy) published before the pilots scaled — and the top support inbound isn’t how to disable ads but “how do I run an ad?”

GPUs are the binding constraint with no end in sight: “demand keeps going up even as prices go down,” tokens-per-user charts are “mindboggling,” and planning “working backwards from GPUs is usually pretty good idea” — humans are hireable and agent-leveraged, “but GPUs are zero sum.”

Code Red is over, exited “which we knew we would” with 5.3 (everyday users) and 5.4 (“workhorse” for knowledge work); the premortem for OpenAI missing its mission “is probably focus,” and the biggest differentiation is the team — “anything we build will get copied.” His long idea: hands-on AI professional services inside real companies, because “we’ve saturated all the emails.”

🔗 Original source & video: ChatGPT – The Super Assistant Era | BG2 Guest Interview

Ali Ghodsi’s central claim: “I think we have AGI. We really have it” — by the definition his 2009 Berkeley AMP Lab used, it’s already satisfied, and the industry is just “moving the goalpost.” He sorts the field into three camps: the superintelligence quest (frontier labs, most of the capital, “I would be very worried there”), the Turing-Award researchers (Sutton, LeCun — sober, 20 years out, “probably the ones that are right, unfortunately”), and camp three — Databricks and Glean — extracting economic value from the AGI we already have.

“The LLM is a commodity” — interchangeable like gas stations, “just compare price,” with users switching models in a day unlike any prior platform battle. Model companies can still be valuable (“TSMC is very valuable”) but as fabs; the real moat is proprietary data and business process — “there’s not an AI out there that understands your secret sauce and your data. That’s not a commodity.”

On the capex math — ~$250B to Nvidia implying ~$500B capex needing ~$1T of AI revenue vs a $400B total software industry — Arvind Jain’s resolution: AI isn’t extending software, it’s converting services dollars, an industry “25 times larger than software.” Ali’s answer is camp-dependent: if superintelligence lands, “any of your cost equations pale in comparison”; camp three doesn’t need it.

Is there a bubble? Yes, but not binary: “there are startups with zero revenue worth 10, 20, 30 billion. That’s a bubble.” Yet both call OpenAI and Anthropic up over 12 months — ChatGPT and Gemini “on fire,” coding having “only eaten into a small portion of that market.”

Ali reframes the MIT 95%-failure stat as healthy: “that’s actually what you want” from an experimentation phase, and hopes for similar stats next year. The working 5%: RBC agents producing equity research notes 15 minutes after an earnings call vs a 2-hour industry standard, Merck’s “Teddy” transformer for gene-regulatory drug discovery, and 7-Eleven’s fully agent-automated marketing stack.

Value accrual call: Arvind thinks the intelligence layer stays thick — “maybe half of enterprise value” — while Ali says most value goes to apps, “I just don’t know which apps”, invoking 1998: everyone bet on Cisco routers and portals, the winners were Facebook/Airbnb/Uber. Software isn’t dead (Salesforce is “a full ecosystem of workflows,” not a database), but data entry is the wedge — “Zoom is really the perfect data entry application.”

Longs and shorts: Ali is long agents and speech (“as long as you’re using a keyboard, we haven’t nailed speech” — keyboards “basically going to disappear”), Arvind calls coding and customer-service automation “a little bit over hyped.” Brad is long proactive AI that comes to the user — the shift that takes “5% power users to 100%.” Glean, fresh off a $200M revenue run rate, is building toward a privileged personal work companion.

🔗 Original source & video: AI Enterprise - Databricks & Glean | BG2 Guest Interview

OpenAI’s enterprise business predates ChatGPT — “the original product for OpenAI actually was not ChatGPT. It was a B2B product. It was the API” — and the company now runs a Palantir-style forward deployed engineer model, embedding at customers like T-Mobile (voice support handled live by OpenAI models), Amgen (a top GPT-5 customer for drug-development paperwork), and Los Alamos (o3 physically installed on the air-gapped Venado supercomputer, shared with Lawrence Livermore and Sandia).

Against the MIT “95% of AI deployments don’t work” headline, Olivier’s pattern from ~200 enterprise deployments: winners have top-down buy-in plus a bottom-up “tiger team,” and evals first — “whenever the customer fails to come up with good evals, it’s a moving target.” Most enterprise knowledge lives “in people’s heads,” not in the SOPs, and hill-climbing from 46% to 99% is “art, sometimes more than science.”

The episode’s sharpest frame: “physical autonomy is ahead of digital autonomy in 2025” despite a higher safety bar — because self-driving had 10-15 years plus roads and stoplights as scaffolding, while “AI agents are just kind of dropped in the middle of nowhere.” Agents date only to the o1-preview reasoning paradigm; “the slope I think is incredibly steep,” and by revenue they may have already crossed Waymo.

GPT-5’s differentiator is “the craft — the style, the tone, the behavior,” not saturated benchmarks: a host said he thought one eval showed hallucinations “basically went to zero,” and the core unresolved trade-off is reasoning tokens vs latency — GPT-5 Pro “one-shots” unsolved problems but takes 10 minutes. The monkey’s-paw lesson: instruction following got so literal that legacy “be concise” prompts broke outputs.

Reinforcement fine-tuning (RFT) — a term Olivier said OpenAI made up — is “an order of magnitude more powerful” than SFT and shifts the pitch from customization to building “a best-in-the-world model” on proprietary data (Rogo in financial services; Accordance hitting SOTA on TaxBench). Olivier’s call: for frontier capability, “RFT will pretty much become the norm.”

The long/short game: Sherwin is short “the entire category of tooling around AI products” — evals products, frameworks, vector stores, and now RL-environment startups — because the stack churns too fast for abstractions to survive a model generation. Olivier is short memorization-based education (“knowledge tokens”) and long healthcare — “probably the industry that will benefit the most from AI.”

Softer signals worth logging: vastly more software engineering ahead even if engineer headcount is ambiguous (“there is a massive software shortage in the world” — OpenAI PMs now ship coded prototypes instead of PRDs); Codex CLI + GPT-5 usage is ripping; and “the blip” (the board coup) left OpenAI antifragile — “a thicker skin and an ability to recover way quicker.”

🔗 Original source & video: Inside OpenAI Enterprise: Forward Deployed Engineering, GPT-5, and More | BG2 Guest Interview

Altimeter’s Apoorv Agrawal calls OpenAI “our largest investment in the history of Altimeter,” entering at last year’s $150B round on super-cycle math. The internet produced Google ($2.5T), mobile produced Apple ($3.5T), and social produced Meta ($1.8T) — so he expects an AI winner to be worth more than $150B, and says “it is clear that ChatGPT has become a verb.” His model: 700M weekly actives and an announced $10B revenue run rate imply roughly $10/user/year on a rough 1B-user basis; over time, 2–4B users at $60–70/user could yield a ~$200B consumer revenue opportunity.

Unlike peers spreading bets across LLMs, Altimeter concentrates because venture’s power law is “not 80/20”: of ~5,000 companies raising yearly, 15 (0.3%) return over 90% of a vintage’s gross profits. His 2001 search analogy: you could have backed Lycos, AltaVista, or Ask Jeeves and been “right” on the charts — but Google took 99% of gross profits by its 2004 IPO. ChatGPT’s user base, he claims, exceeds all rival AI apps combined “multiplied by 10.”

ChatGPT exhibits a retention “smile curve”; the other examples he names are Instagram and TikTok. Time spent now exceeds mainstream apps like X. In the GPT-5 discussion, Molly connected the #4oForever backlash after GPT-4 was turned off to parasocial relationships; Apoorv called it a reminder that “ChatGPT is no longer a website… it is a relationship.” Memory could make switching very hard. He argues it’s still “quite safe and quite healthy” — low dopamine (“no doom scrolling, no cat videos”) and weak network effects, yet dominant anyway.

“Speed is the only moat” — the defensibility is release cadence, not any single model. Operator, Deep Research, ChatGPT Agents, Codex, and GPT-5 all arrived this year by the halfway point; GPT-5’s significance is that it’s the first system, not a model — a router deciding compute allocation per query, “raising both” the ceiling and the floor. On the messy launch, his one-word read on Sam and OpenAI is “antifragile.”

The most tradeable data point: the transcript calls the portfolio company Expo and later also refers to it as XBOW. Its cyber-exploit benchmark success jumped from just under 60% to 81% when switching from Anthropic’s latest model to GPT-5. “The number one hacker in the world is no longer a human”: the company hit #1 on HackerOne in the US within weeks, then #1 globally at Black Hat, with under 50 people spending “more on tokens than we spend on humans.” Apoorv cautions that the cyberhacking ChatGPT moment has not happened yet and says companies like Expo are needed for Western-world safety.

The AI value stack is roughly reversed versus cloud and that’s the biggest open question in AI. Cloud is ~$400B apps / $200B infra / $50B semis; today AI runs at roughly ~$200B semis (NVIDIA alone ~$40B data-center revenue last quarter), $20–30B infra, and $30–40B apps — “90% of all AI dollars are actually in the semis layer.” His analog for patience: AWS started in 2004 and got its first outside customer, Netflix, eight years later.

On the talent wars: Meta generates ~$100B in annualized operating cash flow — more than twice OpenAI’s entire $40B mega-round — so Zuck acquiring talent is “acquiring the most important ingredient, talent, leading indicator of all value.” Compute and data are accessible or saturated; “talent, this is where the war is now.”

Quick-fire: bullish on Klarna, Discord, Databricks, Cerebras, and Anduril — four of five are described as portfolio companies — and he thinks at least one will go public this year. He’d have named Databricks before that morning’s big private round, and concedes that companies such as SpaceX and Stripe may not need to rely on public markets because of repeatable 12–18-month employee liquidity.

🔗 Original source & video: Inside OpenAI’s $500B Valuation | Altimeter’s Largest Bet