Preface Executive Briefing

Week of 29 August 2026

AI Infrastructure: Will Inference Stay Centralised?

Xiaomi put a 3-billion and a 120-billion parameter model in one desktop box.

The footage leaves one important detail unexplained: the rule that decides which model gets the next job. That rule is the enterprise decision.

See Key Takeaways

Explore the briefing

Two models share one box. The switching rule is still private

The hands-on video opens with a small AI-coded flying game, then moves through the Cube hardware and local model demonstrations. The assignment rule remains hidden: the footage never explains how a job reaches the smaller or larger model.

Still from the Xiaomi AI Cube hands-on demonstration — Xiaomi AI Cube hands-on demonstration, 25 August 2026. The original includes Chinese commentary.
Video: 谐门东西 on Bilibili. Watch the original

Engineering prototype

Inside: Three Xring chips and 80GB unified memory
Models: Unnamed 3-billion and 120-billion parameter models
Still unknown: Who or what chooses the model; Xiaomi has not published the switching logic, price or release date.

A contract review and a software test need different paths

The contract contains sensitive information. The software test can run thousands of times and a build can catch a bad answer. Those differences change where the work should run.

Managed frontier model

The provider operates the model and absorbs bursts

Send the request through the provider's API.
Use the strongest model for difficult or unusual work.
Pay as usage rises instead of operating the hardware.

Can this information leave, and does the job need frontier quality?

Company-controlled model

The company keeps the serving path and the operating work

Run downloadable weights on a workstation, appliance or private cloud.
Keep suitable information inside company-controlled systems.
Carry the hardware, power, maintenance and staffing cost.

Does this job happen often enough, and can the local result be checked?

The employee can use one approved interface. The system behind it still needs a rule for every job.

Ask four plain questions before the model sees the job

The answers point to company-controlled capacity, a managed frontier model or a local first attempt with an escalation path.

Can the information leave?

A confidential contract may have to stay inside company-controlled systems.

Can a check catch a mistake?

A software build can reject a bad test before anyone relies on it.

How hard is the job?

An unfamiliar security flaw may need a frontier model and an expert reviewer.

Does it happen often?

Frequent, repeatable work is easier to justify on hardware the company operates.

What the answers change

Keep the job under company control

Use a local model when the information cannot leave and the available model clears the required quality check.

Example: search across confidential contracts

Use the managed frontier model

Use provider capacity when the job is unusually difficult, demand arrives in bursts or the local model misses the quality bar.

Example: investigate an unfamiliar security flaw

Start locally, then escalate

Let a smaller model attempt work that can be checked. Send failures and uncertain cases to stronger capacity and a person when the consequence warrants it.

Example: generate tests, run the build, escalate failures

Turn these questions into thresholds using the company's own work and error tolerance.

A software build can decide when the local attempt is good enough

Routine work can begin on a smaller model because the next step checks it. A failed check sends the same job onward.

The build accepts the test

Stays local

Code changeLocal model writes a testBuild passesDeveloper reviews

The build gives the team an independent check. When it passes, the routine job can stay on company-controlled capacity.

The check changes the destination

The check fails or the answer is uncertain

Escalates

Code changeLocal model attempts the testBuild fails or confidence dropsFrontier model investigatesExpert reviewsreturn to the same workflow

The employee does not choose a provider. The failed check moves the job to stronger capacity, with expert review for a high-consequence case.

This is the mechanism Xiaomi did not show: a visible test decides whether the job stops locally or moves to stronger capacity.

Open models are attracting traffic, and vendors are packaging the hardware

A company can now download a model such as Qwen3.8-27B, test GLM-5.3-Flash through serving software, or buy an appliance. Every option still needs a test against real work.

The local route is moving from experiment to enterprise product

H3C LinSeer MegaCube enterprise AI appliance — H3C markets LinSeer appliances that connect an internal knowledge base and expose local models through an API.
H3C sells this route as an appliance: connect an internal knowledge base, then expose approved APIs to the applications that need them.
H3C product page

23.2T

ox-alpha in six days

9.9T

next model in the chart

Processed tokens show traffic on the route. Answer quality needs a separate test.

A real hybrid system changes what travels

ZTE, Jiangsu Telecom and Jiangsu Guoxin reported that first-stage processing kept raw business data at the enterprise edge. Encrypted intermediate data then travelled to a telecom computing centre for later generation.

Which part of the job needs to leave the site?

The centre stayed in the workflow. Moving the first stage closer to the company changed the information sent there. The partner-reported case shows a split path, while leaving its performance and cost unverified.

Read the ZTE cloud-edge case report

Central capacity is being financed for decades

Local products are becoming more practical while the largest AI companies take on long-lived data-centre obligations. Both can be true at the same time.

Bloomberg network diagram showing investments, hardware purchases and cloud-service relationships among major AI companies — Bloomberg mapped the circular investment, hardware and service relationships around major AI companies.
Bloomberg, Circular Nature of AI Deals

4.25GW

of leased Ohio capacity covered by Nvidia residual-value guarantees

Nvidia's potential obligation is capped at US$105bn under specified conditions.

Nvidia Form 8-K

US$85.2bn

of Alphabet future lease payments, primarily for data centres

The signed leases had not yet commenced; some run for up to 26 years.

Alphabet Form 10-Q

The open question is how much work stays in central data centres.

Local inference already exists. The unresolved questions are how much work moves, how quickly it moves, and whether central capacity finds other demand.

What to do before the infrastructure decision

Start with the jobs, the checks and the consequence of being wrong.

List the jobs before buying the box

Separate confidential work, routine high-volume work and difficult exceptions. Each places different demands on the model and the infrastructure.

Give every local route an exit

A failed test, low confidence or high consequence should send the job to stronger capacity and, where necessary, a person.

Test the vendor headline against real work

Run company documents and real requests through the candidate model. Measure quality, volume, power, maintenance and staff time.

Watch both sides of the capacity bet

Local products are becoming usable while central providers sign long leases and guarantees. The effect on data-centre utilisation is still unproven.

FAQ

The 24 August hands-on video opens with a small flying game made with AI coding, then shows the Cube hardware and local model demonstrations. Xiaomi reported three Xring chips, 80GB of unified memory, and unnamed 3-billion and 120-billion parameter models. It has not published the switching rule, price or release date.