At The Information's "AI Agenda Live" conference held in San Francisco on Wednesday, several speakers said that open-source large models, combined with low-priced new model versions from Anthropic and OpenAI, are helping enterprise customers keep costs under control while AI output stays flat or even improves.

Amjad Masad, CEO of code assistant provider Replit, said: "The existence of open-source models puts pricing pressure on the big model labs, and that's a good thing."

But for many companies, it is still unclear how much value AI investment actually creates. Dan Diasio, global AI leader at consultancy EY, said only one in ten companies can pinpoint in their profit-and-loss statements where the returns on AI investment come from. "Most of the scaling is driven by enthusiasm, not by hard data," he said.

Speakers at the conference came from Google, OpenAI, Blackstone, Uber, and cloud provider Nebius. Below are excerpts from the various sessions.

Uber's Mattie Toia: Uber's AI token spending stays flat

Mattie Toia, Uber's vice president of engineering and infrastructure, said that although internal AI usage has continued to expand, the company managed to keep total AI token spending roughly flat in recent months, with the key being greater efficiency of use.

Earlier this year, Uber's use of AI coding tools led to a budget blowout. The ride-hailing company has since worked to optimize how AI is used and to strictly control costs. Toia said that over the past three to four months, AI costs have stabilized through measures such as memory optimization to eliminate redundant computation and extensive use of Uber's own business data; at the same time, choosing stronger open-source AI models has further reduced spending.

Companies are shifting from "token maximization" — employees using AI without restraint just to show the company embraces new technology — toward what Nebius chief revenue officer Mark Boroditsky calls "value maximization."

Boroditsky said: "What companies really need to pursue is how to optimize the business value created at a given cost." He noted that Nebius is willing to try selling computing power based on the business outcomes customers achieve, rather than billing by NVIDIA GPU hours. Nebius has already piloted spot pricing and auction models for computing resources.

Dion Harris, senior director of high-performance computing and AI infrastructure at NVIDIA, said NVIDIA continues to focus on improving AI operating efficiency, largely through software-layer optimization, such as automatically matching the right model to a task and optimizing memory management. "NVIDIA has always relied on software to add value to hardware platforms, and the open-source Nemotron series of models is one example," Harris said. "A thriving AI ecosystem is good for our overall market expansion, and that is at the core of our strategy, which we will continue to uphold going forward."

Replit CEO Amjad Masad: OpenAI's price cuts slow open-source AI adoption

Amjad Masad, CEO of AI coding startup Replit, said that thanks to OpenAI's cuts to model prices, Replit's use of open-source models in its own products and internal development has decreased compared with January this year.

Earlier this week, OpenAI released GPT-6 Sol and Luna, the second- and third-strongest versions in its latest series, priced 50% lower than models of the same tier in the previous generation. Anthropic simultaneously launched Claude Opus 5.5, which the company says runs at 40% lower cost than its previous top-tier models Fable 5.1 and Mythos 5.1, while matching or exceeding them in coding, reasoning, and other capabilities.

Replit, founded ten years ago, needs to pay model vendors such as OpenAI and Anthropic: when customers call third-party models through its code assistant to generate code, fees are incurred. Replit also uses OpenAI GPT-5.6 Luna to power a free creative brainstorming feature within its product.

But open-source models themselves come with unique challenges. Tamar Yehoshua, chief product and AI officer at collaboration software maker Atlassian, said that although open-source models are highly flexible, deploying and running them creates additional costs, with much of the expense hidden in the internal specialized manpower companies need. "Inference serving itself has hard costs, and running open-source models also requires internal specialists, and that spending is sometimes easily overlooked," Yehoshua said.

—Kevin McLaughlin

OpenAI: Customers have only scratched the surface of AI capabilities

Alexander Embiricos, head of enterprise product at OpenAI, said customers are far from tapping the full frontier capabilities of its models. "This technology is extremely capable, and there is still a long way to go before reaching the upper limit of the value it can unleash. Even if model capabilities no longer iterate as quickly as they do now, there is still a great deal of work to be done at the product level," Embiricos said.

Embiricos said OpenAI refines products based on customer feedback: it typically launches open-ended AI products first, observes which features users actually keep using, and then iterates to lower the barrier to use. The $200-per-month Pro individual subscription plan is a success story for this strategy, with many workplace users relying on it for coding and agentic tasks. "One big advantage of working at OpenAI is that we cover both the consumer market and the enterprise market," he said.

ChatGPT started as a consumer product, but over the past year OpenAI has aggressively expanded its enterprise customer base, coinciding with a surge in popularity for rival Anthropic's coding tools and agents. Embiricos said large enterprises are the first to adopt OpenAI's frontier capabilities, such as coding agents, because their business tasks are highly complex and they have the budget to support advanced models. "We are still at a very early stage: we already have powerful agents, but we still need to simplify operations so people can easily hand tasks off to AI."

Blackstone's Jas Khaira: Blackstone may in the future buy AI lab debt

Jas Khaira, head of Blackstone's AI investment team, said AI infrastructure buildout requires enormous amounts of capital, and Blackstone is still watching which types of capital providers will step in to provide financing. He added that Blackstone is assessing whether debt issued by companies such as OpenAI and Anthropic is attractive as an investment.

Blackstone took part in a partnership to acquire TPU computing assets. Khaira said that due to the computing power shortage, "we plan to expand the scale of procurement." Blackstone's credit and insurance unit is also involved in related investments. Khaira said Blackstone has figured out the market demand for AI computing power, but the types of debt investors suited to such demand are still evolving.

An audience member asked on site: to support infrastructure, must debt yields rise? Khaira responded: "We haven't yet determined who will take on all the debt. We have some guesses, but frankly, the sources of funding are not yet clear."

He roughly divided financing channels into three categories: the public bond market for investment-grade companies, the private credit market, and corporate bonds issued by OpenAI and Anthropic after they go public. "A lot depends on whether they can go public and whether they can subsequently reach investment-grade ratings." Khaira said that at the right price, Blackstone might also acquire AI lab debt. "But we are still watching to see how operations and pricing evolve."

—Dakin Campbell

Block and Cloudflare: AI-driven organizational restructuring brings both opportunities and pitfalls

Executives from Block and Cloudflare said that earlier this year the two companies used AI to drive organizational restructuring, accompanied by large-scale layoffs; they also acknowledged that the reforms encountered many challenges.

In February, Block cut nearly half its workforce, reallocating manpower around AI. The fintech company also built its own tool, Builderbot, which acts as an agent that can fix code defects and independently complete up to 80% of the development work for new features or new products, according to Owen Jennings, Block's business lead.

The change brought a problem: the volume of code exploded, and all of it needs to be reviewed for vulnerabilities, but the company lacks enough code review manpower. "The bottleneck is clearly in code review. Code review is now more important than it was a year ago," Jennings said.

Stephanie Cohen, chief strategy officer at Cloudflare, said the company cut about 20% of its workforce in May. After AI was integrated into workflows, some tasks took far less time: for example, month-end financial presentations used to take nearly a week and now take only an hour.

But companies also need to manage how employees use AI. Cohen noted that after the restructuring, sales team employees all tried to use AI to quickly build similar tools, wasting a lot of time. Cloudflare then set up public training to help employees use AI efficiently.

Surabhhi Gupta, chief technology officer at marketing software maker Klaviyo, believes that employees experimenting with AI tools on their own is a natural evolution of how companies work. Employees come back to work after the weekend and show off: "Look what I made with AI." Companies need to ensure that what engineers build serves customers' core goals.

—Michael Roddan

Neolabs view: To boost enterprise productivity, AI reliability urgently needs improvement

Jerry Tworek, CEO of AI research startup Core Automation and former vice president of research at OpenAI, said AI models are getting stronger in areas such as advanced mathematics and competitive programming, but reliability gains on the vast majority of workplace tasks are limited. As a result, people dare not hand over their jobs entirely to AI agents.

"If AI is so smart, why do we still need to go to work?" Tworek said. "I hope one day I can go on vacation to the beach and let the model do my work. But I realize that to achieve this, we need entirely new algorithms."

Core Automation is researching such new algorithms, exploring architectures more reliable than those of the leading big labs. Top labs have recently achieved a string of impressive results, including OpenAI solving highly difficult math problems, leading some researchers to believe that with existing architectures, AI will eventually surpass humans on all tasks. New-school labs (neolabs) believe existing architectures are still a long way from that goal.

Eric Zelikman, co-founder and CEO of startup Humans &, said coordination among employees, information sharing, and decision-making ability limit how effectively companies can deploy AI products. Humans &, founded just one year ago, aims to build AI that helps users achieve long-term goals. Zelikman said companies such as OpenAI and Anthropic are in a "desperate race" to automate as many individual tasks as possible (such as coding), while Humans & takes a different path.

One notable feature of Humans &: most of the hardware used to train its AI is self-owned. Zelikman estimates that renting two sets of AI chip clusters would cost more than $1 billion; the company is deploying a third cluster.

—Rocket Drew

How leading AI labs earn revenue several times their annual computing costs

OpenAI and Anthropic are racing to lock in trillion-dollar-plus computing contracts while preparing for IPOs. The market has never had a clear answer on how much revenue these huge orders will actually generate.

One data point suggests there is room for returns on this investment. Ricard Boada, co-founder and CEO of Volta, a provider of data center financing, energy, and computing capital services, said frontier AI labs currently monetize $75 million per megawatt of computing power per year, while the levelized cost of that computing power is only $15 million to $20 million.

Boada explained that leading labs have market pricing power, and potential returns can be several times operating costs, which is why they are eager to deploy power and chips at scale. Volta emerged from stealth mode in August this year, announcing a $10 billion computing power deal with a frontier AI lab in Norway, with reports saying the customer is Anthropic. Boada declined to confirm the customer's identity, saying only that the contract includes a $5 billion hardware purchase order and a $2 billion data center contract for which Volta needs to arrange financing.

Chris Dolan, chief data center officer at Crusoe, said at the same forum that building a 1-gigawatt data center can cost up to $15 billion to $20 billion, with an additional $30 billion to $40 billion needed for chip procurement. Top frontier labs could theoretically recoup their investment quickly, but only if power delivery and computing power coming online are highly coordinated.

Recent market transactions show that public equity and bond investors may be less willing to finance AI infrastructure. Boada and Dolan believe that as the industry matures, AI data center contracts can deliver predictable returns, but capital providers are becoming increasingly strict in selecting investment targets. "Capital markets are adapting to this entirely new asset class," Dolan said.