AT&T runs 40% of AI workloads on open models, targeting 70%

The Financial Times reports that large companies outside the AI industry are moving real production workloads onto open-weight models: AT&T says 40% of its AI workloads already run on them and it is targeting 70% within a year, and Tinder's AI spend rate grew from $1 million to $10 million a year in six months, with non-technical queries routed to open weights.

Key facts

• 40% now, 70% targeted — AT&T chief data officer Andy Markus told the FT that about 40% of the company's AI workloads run on open-weight models, targeting roughly 70% within a year, at a reported 45 billion tokens per day.
• $1M to $10M in six months — Tinder CTO Vinay Kuruvila said the company's AI spend rate grew from $1 million a year in January to $10 million a year by July, with non-technical-user queries routed to open-weight models.
• "Never, never, never" into a frontier model — Digital Realty's Scott Wallace on customer data, citing sovereignty and security as the driver for keeping prompts inside the company's own environment.
• 56% open-weight token share in August — Vercel's AI Gateway Production Index, up from 7% in December, with a 78.4% single-day open-weight share on September 19.
• A sixfold surge in executive mentions — AlphaSense recorded six times more mentions of open-weight or open-source AI in executive communications in August–September versus the same period last year.

What happened

The Financial Times' report, published over the September 26–27 weekend, is the most detailed look yet at open-weight adoption inside companies that buy AI rather than build it. The named adopters are not startups. AT&T is a telecom carrier processing AI workloads at a scale its chief data officer put at 45 billion tokens per day. PNC Financial Services, CH Robinson and Siemens also appear in the piece as adopters.

The reasons given split into two buckets, and the FT is careful to keep them distinct. Cost is the one executives quote most readily. Tinder's Kuruvila told the FT the company routes queries from non-technical users to open-weight models and keeps frontier models for harder requests, adding, "if open-weights models catch up, I may not need to use [frontier] anymore." His spend numbers — a tenfold increase in six months, driven largely by volume — explain the pressure to arbitrage between model tiers: at 45 billion tokens a day, AT&T's potential savings from even a per-token discount are measured in millions.

Sovereignty is the quieter driver, and possibly the more durable one. Scott Wallace of datacenter operator Digital Realty told the FT, "any customer data, never, never, never goes into a frontier" model. For companies under regulatory or contractual data-handling constraints, the choice is not open versus closed on quality; it is whether prompts leave the building at all. Self-hosted open weights are the only option in that frame.

The FT's macro framing figures — that open weights account for more than half of token usage but only about 14% of spend, and its reporting on Anthropic's $2 trillion IPO expectation and OpenAI's $1.2 trillion private-round talks — are the FT's reporting and analysis, and we present them as such. The token-share claim is at least partially checkable against independent data: Vercel's AI Gateway Production Index recorded 56% open-weight token share in August (up from 7% in December) and a 78.4% single-day open-weight share on September 19. We covered that data when it published, and it corroborates the direction of the FT's token figure, though Vercel's index reflects one gateway's traffic, not the market as a whole. The 14%-of-spend figure has no independent corroboration.

Why it matters

For a builder deciding where to run models, this piece is a decision framework in source form, expressed by people with budgets attached. The three questions the FT's adopters answer are the same ones that show up in our self-hosting versus API cost breakdown: at what token volume does the switch pay, what do you keep on frontier models (Tinder keeps hard queries there), and where does data sensitivity force self-hosting regardless of price.

The spend-side math is the part most often glossed over. Tinder's numbers show AI budgets growing faster than most engineering line items — tenfold in six months — which is exactly the situation where per-million-token price differences stop being rounding errors. Routing by query difficulty, as Tinder does, is the cheapest optimization available: it requires no infrastructure, only a classifier and a fallback. The current open-model price index shows capable open-weight models priced well below frontier tiers, and models like DeepSeek V4 Flash are what "route the easy queries" looks like in practice.

The sovereignty bucket is the one that doesn't show up in price comparisons at all. A company whose policy is "customer data never enters a frontier API" is not shopping on price; it is shopping on deployment model. That is the demand that keeps self-hosting and private-inference infrastructure growing even when frontier APIs are cheaper per token, and it is why the open-source API landscape now spans everything from bare metal to sovereign deployments.

Background

The FT piece is mainstream-business confirmation of a shift the trade data showed earlier. Vercel's gateway data — 56% open-weight token share in August, against 7% in December — came from a developer-tools audience, the population most likely to experiment first. The FT's subjects are the second wave: banks, logistics firms, telecom carriers, industrial companies. When the adoption curve reaches companies whose core product is not software, the experiment phase is over.

The capacity side has kept pace. The current generation of open-weight releases — mixture-of-experts models with per-token costs a fraction of frontier pricing — is what makes a 70% target credible rather than aspirational. A year ago, routing 70% of a telecom's AI workloads to open weights would have meant a visible quality drop on most tasks. The capability gap that remains is concentrated in the hardest queries, which is precisely the fraction Tinder reports keeping on frontier models.

AlphaSense's sixfold increase in executive mentions matters as a leading indicator. Executive communications lead procurement by quarters, not days. If the mention curve continues, the 2027 question will not be whether enterprises adopt open weights but how they audit them — which shifts attention toward verification, evaluation, and the operational discipline of running models you patch yourself.

What's next

Watch three things. First, whether AT&T's 70% target becomes a stated deadline with a published architecture, which would make it the reference enterprise deployment for this cycle. Second, whether the FT's ~14%-of-spend figure gets independent corroboration — if open weights really are capturing tokens far faster than revenue, frontier API pricing pressure follows. Third, whether Tinder's routing pattern (open weights for volume, frontier for difficulty) shows up in other companies' published architectures. The next data point arrives with Vercel's October index: if open-weight token share holds above 56% through a full month of enterprise traffic, the August number was a floor, not a spike.