Goldman Sachs says open-source AI could be Big Tech’s unexpected winner as AT&T cuts model costs by 56%

Open-source AI was supposed to squeeze the economics of the industry’s biggest players. Goldman Sachs sees another possibility: cheaper models could make enterprise AI affordable enough to drive far more usage, helping Big Tech fill the enormous computing capacity it is spending billions to build.

AT&T may already be putting that thesis to work.

The telecom giant has cut the cost of coding and some other advanced AI tasks by as much as 56% through tools that route employee queries to cheaper models when the job does not require a premium frontier model, The Information reported Thursday.

AT&T sends simpler tasks, such as document and code summaries, to open models including Meta’s Llama and Google’s Gemma. Its developers can still use more advanced models from companies such as Anthropic and OpenAI for demanding work, including complex code generation.

The approach produced a striking result: costs for some coding tasks fell 56% with just a 2% decline in quality, according to The Information.

That matters far beyond AT&T. It points to a change in how companies may buy and use AI. Instead of asking which model is best, enterprises are starting to ask a more economically useful question: What is the cheapest model that can reliably handle this particular job?

Goldman Sachs sees cheaper AI driving more AI

Jim Covello, head of Goldman Sachs Equity Research, believes advances in smaller, cheaper open-source models could benefit hyperscalers such as Amazon, Microsoft and Google rather than simply threatening their AI businesses.

“Open-source models will make it more likely that enterprises can profitably implement AI in the organization,” Covello said during a recent Goldman Sachs Wealth Management Investment Strategy Group discussion.

The logic is counterintuitive. Open models can reduce the price companies pay for individual AI workloads, putting pressure on premium model economics. Lower prices can also make far more AI applications financially practical.

More viable applications could mean more queries, more tokens processed, and greater demand for the data centers and computing infrastructure sitting beneath those models.

“I think it’s really good for the hyperscalers, because it’s more likely that you’re going to be able to profitably fill up all this capacity that you’re adding,” Covello said.

Watch the full video for Covello’s views on which types of companies are likely to profit from AI.

That capacity has become one of Wall Street’s biggest questions.

Big Tech companies have committed vast sums to AI chips, data centers, networking and energy infrastructure. Investors initially cheered rising capital expenditures as evidence that companies were positioning themselves for the AI boom. Covello says that mood has started changing.

“Until relatively recently, every time one of these companies would announce higher capital expenditures, the market would reward that company. Over the last quarter or so, you’ve really seen a significant shift where the market is questioning that a lot more,” he said.

The question is no longer whether Big Tech can build enough AI infrastructure. It is whether companies can generate enough profitable AI demand to justify what has been built.

Cheaper open models could help close that gap.

AT&T offers an early look at the model-routing economy

AT&T is already processing about 45 billion AI tokens per day, according to The Information. Roughly 40% of its employee AI queries are being routed to open-source models, and the company expects that figure to reach 60% to 70%.

Mark Austin, an AT&T vice president, told The Information that he has found open-source models are “just as good or better” than older models sold by companies including Anthropic and OpenAI for many jobs.

“For instance, AT&T’s software developers still rely on cutting-edge models for complex tasks like generating code, but can use cheaper open-source models for less intense tasks like generating summaries of previously submitted code,” The Information reported.

“We expect that to just keep getting better going forward,” Austin said.

This is where AT&T’s experience intersects closely with Covello’s argument.

Covello identifies what he calls the “model optimization layer” as a technology bottleneck for enterprise AI. The idea is straightforward: high-consequence requests get sent to expensive frontier models, and routine requests go to cheaper open models.

“That’s going to be one of the big keys to unlocking the economic value of AI in the enterprise,” Covello said.

If that model takes hold, the AI competition could start looking very different.

Enterprises may use several models rather than committing most workloads to one provider. Model routers could make decisions based on cost, quality, latency and task difficulty, treating AI models increasingly like interchangeable computing resources for routine work.

That could put pricing pressure on frontier model companies. Yet it could create a much larger market for AI inference.

AT&T is a useful early test. The company reportedly expects greater use of open models to help keep spending on premium AI services flat as its total AI usage grows.

That is the paradox at the center of Goldman’s thesis: cheaper models could mean less revenue per query, but vastly more queries worth running.

For hyperscalers spending billions to build the infrastructure underneath those queries, open-source AI may prove less of a threat than an unexpected source of demand.