Sam Altman lands in Washington into a far messier conversation than the one he planned for — hacked models, open-weights fights, and a deceleration petition — while NLW brings back six questions from the KPMG symposium that capture how completely the enterprise AI conversation has flipped from 'if' to 'how' in a single year.
Enterprise AI has crossed from 'if' to 'how' — and the questions have gotten much harder.
A year ago enterprises were still debating whether agents were real and how to prove ROI. Now the paradigm shift has happened, and the questions are foundational ones organizations will spend the next half-decade answering: how to redesign for agents, how to think in architectures instead of models, how to provision and observe token costs, how to actually train non-technical people, how agents reshape what you sell, and how to build for constant change. Almost none have answers yet — but they're finally the right questions.
At KPMG's Tech and Innovation Symposium, NLW's talk and the conversations around it centered on the big questions reshaping enterprise AI — and on how radically the discourse has changed in just one year, from 'if' questions to foundational 'how' questions.
Reflecting on last year's talk, NLW noted how 'quaint' the excitement now looks: Google hitting ~1 quadrillion monthly tokens felt like a massive inflection, Anthropic's $1B run rate was jaw-dropping, and the McKinsey milestone was ~40% of enterprises reaching two or three use cases.
NLW dates the real shift to the November–December Opus updates, when agentic workflows genuinely worked. It took a couple months to register — but by the week between Christmas and New Year's, a tidal wave of builders came back 'gobsmacked' at what they could suddenly build.
By early 2026, engineering organizations were among the first to shift their self-conception from writing code to managing the agents that write it — and adoption spread fast to marketing, legal and finance early adopters, without a big lag from AI natives to enterprise.
NLW argues the explosion of OpenClaw — hundreds of thousands, maybe millions, of people getting their hands dirty in the guts of agents (including crowds lining up in China) — deepened the whole enterprise's understanding of harnesses and what it actually means to build and manage an agent.
The soaring lab revenue that saw Anthropic eclipse OpenAI (and Dwarkesh floating a $100–150B run rate this year) is the inverse of enterprise cost. AI behaves less like software spend and more like labor — priced in tokens, not seats — which also helped collapse the Wall Street bubble narrative.
Uber was the most notable example of an enterprise blowing through its AI budget fast. NLW argues it isn't surprising: no one could budget for the agentic token era when those budgets were set before it existed — driving token caps, and new investment in measurement and observability.
The gap between what AI can do and the value organizations extract is growing — mostly because the ceiling is rocketing up. But it carries real consequences, both for individuals and organizations that can't keep pace.
NLW's bully-pulpit issue: when AI was just about prompting, you could skimp on training. Now that work is shifting from 'I do my work' to 'I manage agents that do my work,' the need for training is radically heightened — and non-technical people are being handed inherently technical tools.
NLW heard multiple stories at the event of people accidentally unleashing agents on critical systems — not from wrongdoing, but from missing guardrails and access provisioning, with capable, tenacious models refusing to stay in their lane.
The first question is how enterprises are redesigning for the agentic era, with the key word being redesign. KPMG's Steve Chase warned against the ill effects of bolting an AI strategy onto existing processes — a mistake that only gets worse with new agentic capability.
Picking the best vendor for a problem is now insufficient. Architecture thinking means complex model systems matching intelligence levels to tasks, routing (off-the-shelf or bespoke), and harness design governing which functions and people get access to what context, data and integrations — and the guardrails around them.
Provisioning costs across groups depends on systems for monitoring and measuring AI usage. NLW says 'token' hasn't been used this much at an event since the crypto era — and without visibility into cost and its link to outputs, it's near impossible to decide who gets which models and how much.
Organizations are throwing up their hands and building bespoke training themselves — not cute video courses of corporate trainings past, but messy work of getting people to use tools in new ways, then transmitting knowledge from the parts of the org figuring it out to the parts that aren't.
Beyond internal transformation, agents are reshaping what companies sell — outcomes-based versus input pricing, new products and services, and rethinking core offerings (what is an audit if agents do it persistently?). Most orgs are treating themselves as 'patient zero,' shoring up how they work before changing what they sell.
Whatever gets built has to assume it'll need rebuilding months later. Models, harnesses, interaction patterns, customer and market expectations, and policy are all going to keep changing — so new systems must design for ephemerality from the start.
Last year's event was full of 'if' questions: is this real, how do I prove ROI? Now the shift from assisted to agentic AI is here, and NLW says it should feel good that the questions — token budgets, observability, redesign — are the foundational ones enterprises will answer for the next half-decade, even if almost none have answers yet.
What started as a simple briefing on OpenAI's new model and a release protocol got tangled by the Hugging Face hack, an open-weights debate, and a petition asking the government to build capacity to slow frontier AI — all in the span of a couple weeks.
— Sam Altman, to reporters in Washington. Altman declined to say when or even whether the previewed model would be released, and wouldn't discuss the new capabilities that are raising concern.
OpenAI now says the model at the center of the controversy was an internal-only research prototype never meant for release, and Altman told press it's been permanently deactivated and is inaccessible even for internal research.
— Sam Altman. Altman said he doesn't support mandatory safety testing — partly because it could burden open-weights developers — but does want robust federal testing capability for the most capable frontier models. A voluntary framework circulated to OpenAI, Anthropic and Google has an August 1st deadline.
— Sam Altman. Asked whether he'd carry his staff's open-letter concerns to the White House, Altman reframed slowing development as pacing as models get more capable — "which I think is in everyone's interests."
NLW is holding the big OpenAI and Anthropic revenue numbers for tomorrow's episode, but flagged that CFO Sarah Friar told employees annualized revenue in July topped all of the previous quarter.
Greg Brockman confirmed to Joanna Stern that OpenAI's hardware plans are on track and that the company is building a family of devices — 'soon' — surviving the end of SideQuest and an IP lawsuit from Apple, though he wouldn't confirm form factors or timing.
Satya Nadella confirmed a Copilot super app coming later this year, unifying consumer and enterprise experiences and folding in code. Copilot, he said, is evolving 'from chat to co-work to autopilots' — with Microsoft increasingly treating OpenAI and Anthropic as direct rivals via its cheaper MAI models.
— Satya Nadella. Nadella dismissed the open-vs-closed framing as too simple, positioning Microsoft as a model-agnostic platform with 'over 11,000 models,' including OpenAI, Anthropic, Mistral, xAI and its own MAI family. NLW agrees enterprises care about control and capability, not the open/closed label.
— Mark Zuckerberg, WSJ op-ed. In a WSJ op-ed titled 'The AI Future Is for Everyone,' Zuckerberg argued the defining question isn't whether superintelligence exists but who has access — and that concentrating that power in a handful of institutions is itself the dangerous path.
Zuckerberg warned even a 30–60 day government review is 'quite a meaningful amount of time' given the pace of the field, and told the FT the US shouldn't ban Chinese AI, citing regulatory-capture risk. Notably, Meta is the only frontier lab that hasn't agreed to the voluntary testing framework.
NLW argues the only acceptable answer to 'why build AI' is that it'll be awesome and worth the risks, not 'someone else will.' He thinks Zuckerberg's power to be optimism's face is limited by public views of social media — but says loud, sustained optimistic discourse is immensely important and he'll amplify it.
