Average AI inference prices hit their lowest point of 2026 in early August, down from $2.04 per million tokens at the end of May to around $1.17. Open-weight models, mostly from Chinese labs, went from 4% of tokens on one major gateway in January to 55% in July. On the Bricks & Bytes round table, Dustin DeVan’s observation was that almost nobody in corporate America has right-sized which model does which job. You do not need frontier reasoning to search an inbox, and paying for it anyway is quietly funding the data center boom.
Owen opened this segment with the framing the whole episode had been building toward. What is happening to the price of AI, why is everyone moving to Chinese models, and what did Meta just do?
Martin’s instinct was that cheaper is good news. Owen’s correction was that it depends entirely on who you are. If you are buying tokens, a price collapse is a gift. If you are a foundation lab that has committed to enormous capital expenditure on the assumption of enormous future revenue, it is something else.
Dustin had been with a very large national builder days earlier and came back with the observation that stuck. They had not right-sized which models they were using. That is the state of most enterprises right now, and it has consequences that run all the way back to the construction backlog.
The numbers behind the collapse
Prices roughly halved in ten weeks
Jefferies, citing data from research firm Silicon Data, reported average inference prices between $1.16 and $1.18 per million tokens across 6 to 8 August. That was the lowest level recorded this year, down from $2.04 on 31 May and $1.45 in late July. The index tracks pricing across business API providers and open-weight inference platforms.
The pressure is coming from both directions. OpenAI cut rates on its latest model series by up to 80%. Anthropic’s Claude Opus 5 delivered performance comparable to its flagship at half the price, according to the same Jefferies work. Meanwhile open-weight releases have been arriving at a pace that makes quarterly procurement reviews look quaint.
The share shift is the part that should get a CIO’s attention. Open-weight models reached 55% of token volume on Vercel’s AI Gateway in July 2026, up from 4% in January. Separately, CNBC reported that US companies routed more than 30% of their OpenRouter tokens to Chinese open models every week from February, touching 46%, against a twelve-month average nearer 11%. Estimates of the cost gap sit at 60% to 90% cheaper than closed frontier models for comparable work.
Right now people use bazookas for something you want a hand knife to do.Dustin DeVan, founder of Ediphi, on the Bricks & Bytes round table
Right-sizing is the whole game
Nobody needs frontier reasoning to sort a calendar
Dustin’s framing of the enterprise question was refreshingly practical. Do you need the latest model from a frontier lab for someone searching emails, or doing calendaring and scheduling? Probably not. What you need is whatever is most cost efficient for that specific job in your organization.
Martin’s counter was about where the difference actually bites. If the gap between the top model and the tier below is one and a half times rather than a hundred times, the calculus changes for most work. Nobody building rockets is going to run an old model, and that is fine, because they are in a competitive position where paying top dollar is rational. Most businesses are not.
Dustin sharpened it around urgency rather than regulation. If your engineer’s time is genuinely valuable and they are writing code, use the best model available. If you are an admin processing invoices, you are not in the same economic situation. His view is that the step change comes when models can actually reduce headcount, at which point paying frontier prices justifies itself. Corporate America is realizing that is not happening yet, which is the same argument he made about construction productivity earlier in the episode.
Owen raised the honest counterargument. Open-weight models are typically built on slightly older versions of the most advanced systems, so you give up access to the frontier layer. That is a real trade, and the answer for most organizations is routing rather than a wholesale switch.
The AI takes construction execs actually forward to each other
What is working, what is priced wrong, and what is still five years away. Weekly.
Join 3000+ ReadersThe construction-specific catch
Cheap tokens and sensitive project data are an awkward pairing
Owen raised the issue that makes this more than a procurement conversation for AEC firms. American companies working on potentially sensitive projects are now processing information through open-weight models, and depending on where those models run and who built them, that becomes a security question rather than a cost question.
It is worth being precise about the distinction, because the show was not, and Patric said as much. Open source, in the traditional sense, means the code is visible, modifiable, and redistributable. Open weights means the trained model parameters are downloadable, so you can run and fine-tune the model yourself without necessarily seeing how it was built. Patric attempted the distinction on air and immediately flagged that an AI researcher would probably wince at his version. Credit for the honesty.
The practical consequence for construction is the useful bit. Open weights means you can host the model on your own infrastructure, in your own jurisdiction, fine-tune on proprietary project data, and pin a version that will not be deprecated underneath you. That is why regulated buyers who would never send data to a foreign API endpoint are still willing to run those weights on their own hardware. The risk profile of the model provider and the risk profile of the deployment are different things, and treating them as one is how firms end up either paranoid or exposed.
If your organization is still working out who owns this decision, the governance gap is the actual problem rather than the model choice. We looked at how specialist practices are trying to close it in KP Reddy Co. Launches Embedded AI Transformation Practice for AEC Firms.
Why this loops back to your backlog
If revenue diverts to cheaper options, the build slows
Dustin closed the loop in a way that should interest anyone with data center work on the books. The foundation labs are spending enormous sums building out capacity because they expect enormous revenue. If a meaningful share of that revenue diverts to much cheaper options that are a better fit for the actual utilization case, those labs may slow their own spending as they recognize the revenue curve is flatter than modeled.
That is a demand-side risk to the build cycle that has nothing to do with construction capability and everything to do with token pricing. It sits alongside the financing questions we covered in Big Tech Promised $650 Billion in Data Centers, Most of It Isn’t Being Built, where announced capacity and capacity actually under construction have been diverging for months.
Martin’s contribution was about market structure. His reading is that if only one or two providers were permitted to serve the market, the trillion-dollar revenue scenarios become achievable. With genuine competition and no regulatory moat, competition sets the price of what is becoming a commodity for most use cases. Whether inference is genuinely commoditizing or simply repricing is the open question, and the answer determines a lot of construction backlog three years out.
| Task | Model tier that fits | Why | Deployment note |
|---|---|---|---|
| Complex code, deep research, agentic engineering | Frontier closed models | Capability gap still matters when the operator’s time is expensive | Vendor API, standard commercial terms |
| Contract and specification review | Mid-tier or fine-tuned open weight | Domain accuracy beats raw reasoning, and volume is high | Consider self-hosting where documents carry liability |
| Email search, scheduling, meeting notes | Cheapest capable tier | No frontier reasoning required, and volume compounds cost fast | Vendor API is usually fine |
| Invoice and document processing | Open weight or small commercial model | Structured, repetitive, well-bounded | Self-host if volumes justify the operations overhead |
| Sensitive or security-cleared project data | Open weight, self-hosted | Control over jurisdiction and data residency is the requirement | Run on your own infrastructure, audit the deployment |
Jefferies, citing Silicon Data, reported average inference prices of $1.16 to $1.18 per million tokens between 6 and 8 August 2026, the lowest recorded this year. That compares with $2.04 on 31 May and $1.45 in late July. The index tracks pricing across business API providers and open-weight inference platforms used by developers. Jefferies analysts attributed the decline to an increasing emphasis on cost efficiency across both the US and Chinese ecosystems, alongside direct price cuts from major providers. (Source)
Open source in the traditional software sense means the source code is available to inspect, modify, and redistribute. Open weights means the trained model parameters are downloadable, so you can run the model yourself, fine-tune it on your own data, and host it wherever you choose, without the training code or data necessarily being disclosed. The two get used interchangeably and they are not the same thing. Patric attempted the distinction on the show and flagged his own answer as probably only half correct, which is a fair description of how most of the industry uses the terms.
On the gateways where it can be measured, yes. Vercel’s data showed open-weight models reaching 55% of token volume through its AI Gateway in July 2026, up from 4% in January, with DeepSeek alone commanding 22.6% of tokens through June. CNBC reported US companies routing more than 30% of OpenRouter tokens to Chinese open models weekly from February, peaking around 46%, against a twelve-month average nearer 11%. These are developer routing platforms rather than the whole market, so read them as a leading indicator rather than total enterprise share. (Source)
Routing beats switching. The consistent recommendation across analysts covering this shift is to match model tier to task rather than standardize on one provider. Keep frontier capability for the small number of workflows where reasoning quality changes the outcome and the user’s time is expensive. Move high-volume, well-bounded work to cheaper tiers. Hard-coding an application to a single provider has become a commercial liability given how fast relative pricing and capability are moving. A monthly benchmarking review against your own workloads is now a reasonable operating cadence. (Source)
Owen raised this on the show in the context of American firms working on projects where the information itself is sensitive. The key distinction is between sending data to a third-party API endpoint and running downloadable weights on infrastructure you control. Open weights allow self-hosting in your own jurisdiction, version pinning, and auditing, which is why some regulated buyers who would not use a foreign-hosted API will still run those weights internally. Any organization handling defense, critical infrastructure, or client-confidential project data should treat model provenance and deployment location as separate decisions requiring separate sign-off.
That was Dustin’s argument on the show, and it is a coherent one. Foundation labs are committing to very large capital expenditure on the expectation of very large revenue. If enterprise workloads migrate to cheaper options because those options fit the actual use case, the projected revenue curve flattens, and the labs may moderate their own build plans in response. This is a demand-side risk that sits alongside the supply-side constraints already slowing projects, including transformer lead times and grid connection queues. It is a scenario rather than a forecast. (Source)
Meta announced a national partnership with North America’s Building Trades Unions in August 2026, working with NABTU to develop registered apprenticeship programs for data center construction. It builds on America’s Workforce Academy, a $115 million first-year initiative Meta launched in June 2026 with the Associated Builders and Contractors and CBRE. NABTU represents more than 3.2 million skilled craft professionals across 14 unions and operates more than 1,900 apprenticeship facilities. Dustin’s take on the show was approving, noting Meta is at least putting money to work where others are not. (Source)
Related Articles
KP Reddy Co. Launches Embedded AI Transformation Practice for AEC Firms
Big Tech Promised $650 Billion in Data Centers. Most of It Isn’t Being Built
Construction’s Coordination Problem Is Now a Budget Problem
SCMP: enterprise AI costs hit a 2026 low
Chinese open models and enterprise AI cost, 2026
Scott Logic: the rise of open weights
Meta and NABTU partnership announcement
Construction Dive on the Meta and NABTU deal
