Why outsourced and local AI is closing the gap on frontier labs for business

For the past three years, the AI conversation in boardrooms has been remarkably one-sided. If you wanted serious capability, you went to a frontier lab — OpenAI, Anthropic, or Google — paid the per-token rate, and accepted that your data, your latency, and your roadmap were tied to a handful of US west-coast companies. That bargain is starting to look less obvious. A growing chorus of analysts, including Ben Thompson at Stratechery, have argued that the economics are quietly shifting toward a different stack: capable open-weight models, deployed locally or through specialist outsourcing partners, at a fraction of the cost.

For Australian businesses — many of whom are wary of cross-border data flows, currency exposure on USD-denominated API bills, and the latency penalty of round-tripping to North American data centres — that shift matters. This is an attempt to unpack why it's happening, where the savings actually come from, and what the trade-offs look like.

The frontier premium is real — and shrinking

Frontier labs charge what amounts to a research premium. The price of a token from a top-tier closed model has to cover not just inference costs but the multi-billion-dollar training runs, the safety teams, the data licensing, and the lab's race to the next generation. As long as those models were materially better than anything else available, customers paid willingly.

What's changed is that the gap between the best closed model and the best openly available model has compressed dramatically. Open-weight releases from Meta, Mistral, DeepSeek, and others now sit close enough to the frontier on most business-relevant benchmarks — summarisation, classification, retrieval-augmented generation, code completion — that the marginal capability difference no longer justifies a 10x or 20x cost gap for routine workloads. Thompson and other observers have repeatedly made the point that strategy and impact in technology usually turn on commoditisation curves, and AI inference is following one.

The frontier still leads on the hardest, most novel tasks. But most enterprise AI use isn't novel. It's tagging support tickets, drafting emails, extracting fields from PDFs, and answering policy questions over an internal corpus. For that, you don't need GPT-class reasoning; you need a competent model that runs cheaply and predictably.

Where the cost savings actually come from

When people say "local AI is cheaper," they often blur three different things. It's worth separating them:

The interesting move in 2024 and 2025 has been the emergence of a middle layer: outsourcing firms — often the same systems integrators that once sold SAP rollouts or managed AWS estates — that will stand up an open-weight model for you, fine-tune it on your data, host it in a jurisdiction you choose, and bill on a predictable basis. That's the combination that changes the maths. You get someone else's operational expertise (you don't need to hire a research team), but without the frontier API's per-token economics.

What this looks like for Australian businesses

For mid-sized Australian firms, the calculus is becoming concrete. A bank or insurer with strict APRA expectations around data residency can now plausibly run a capable model inside its own cloud tenancy in Sydney, rather than negotiating data-handling agreements with an overseas lab. A mining services company in WA can fine-tune a model on equipment manuals and field reports without exposing that corpus to a third party's training pipeline. A government department can specify that inference never leaves Australian soil — and actually mean it.

The outsourcing angle matters here because most Australian organisations don't have, and won't acquire, the talent to operate this stack themselves. The supply of ML engineers is thin and expensive. What's becoming available instead is a services market: consultancies and specialist providers who handle the infrastructure, model selection, evaluation harnesses, and ongoing tuning, and who can do it across multiple clients to spread the fixed costs.

That's a familiar pattern. It's how mid-market companies eventually consumed mainframes, then cloud, then enterprise SaaS. AI is just travelling the same path faster.

The trade-offs to be honest about

None of this means frontier APIs are going away or that local deployment is automatically the right call. There are real reasons to keep paying the premium:

The honest framing is that for a portfolio of AI workloads, the answer is increasingly hybrid. Route the hard, novel, low-volume queries to a frontier API. Run the high-volume, well-understood workloads on a local or outsourced open-weight deployment. The savings come from refusing to pay frontier prices for non-frontier work.

What to do about it now

If you're a decision-maker watching this shift, three practical moves are worth considering this year. First, audit your current AI spend by workload, not by vendor — you almost certainly have a long tail of routine tasks that don't need a frontier model. Second, build or buy an evaluation harness so you can actually measure whether a cheaper model meets your quality bar; without that, every "we tried open-weight and it wasn't as good" claim is just vibes. Third, talk to two or three outsourcing partners who can demonstrate working deployments, not slideware — the gap between providers in this space is enormous.

The broader point, which Thompson and others have been making for some time, is that the AI market is following the same commoditisation logic as every previous wave of computing. The frontier moves forward, but yesterday's frontier becomes today's commodity, and the economic surplus migrates from the labs that built the technology to the businesses that deploy it well. For Australian organisations that have spent two years feeling like AI tourists in someone else's economy, that's a more hopeful place to be.

Related on Bleen

Sources