The AI compute boom just ran straight into export-control enforcement. New to this story? Here's where things stand. Demand from model labs and enterprises is driving big physical buildouts at specialist AI-infrastructure providers. CoreWeave, Together AI, Cerebras, Nebius, and Bitdeer have all reported new capacity, backlog, or financing milestones. Bitdeer AI, for example, recently secured a five-year contract worth roughly $400 million for about half of its 9.5MW A102 AI data center in Malaysia—before the facility is energized. This is AI Daily Briefing. A giant Georgia compute deal, alleged server smuggling, and a fight over who controls open-source inference—let's start with the contract where the customer kept an escape hatch. From The Economic Times:
Under the deal, the customer will get access to GPUs and related services from RUM Group's Maysville, Georgia site, which is currently under development. The customer will buy $13.7 billion in GPU services over six years in three parts, with the last instalment contingent on a delivery date approved by the client.
The neocloud backlog story is getting bigger: RUM Group says an unnamed U.S. cloud customer signed for $13.7 billion in GPU services. It's a six-year deal at Maysville, Georgia—except Maysville is still under development. And tranche three only comes due if the client approves the delivery date. Good—construction risk belongs with the builder. If that site slips, $13.7 billion becomes a very expensive press release with two completed payments. RUM—formerly Rumble, the Truth Social host—also gave this unnamed buyer options on roughly 51 million shares at one cent apiece. So the customer gets GPUs if Georgia comes online, plus equity upside while it waits. The eight-percent premarket pop is the market buying the headline. I'm watching the client-approved date. Until Maysville is delivering compute, the contingency clause is the only hard thing here. Ars Technica writes:
Taiwan has indicted nine people— reportedly including an Nvidia senior manager and two Supermicro employees based in the country—who allegedly helped forge documents to cover up illegal exports of high-end AI servers to China in violation of US export controls. On Monday, Reuters reported that prosecutors in Keelung did not release any names but confirmed that the suspects were “charged with breach of trust and document forgery.”
Nine indictments over forged export documents, with a reported Nvidia senior manager and two Supermicro employees in the mix. If the allegations hold, internal controls failed right across the server supply chain. Keelung prosecutors are charging breach of trust and document forgery—not issuing another policy memo. We just covered a $13.7 billion infrastructure contract; this is the other side of the same hardware business, where enforcement lands in the shipment paperwork. The U.S. license rule has been in place since 2022. It only works if everybody touching the order—Nvidia, Supermicro, distributors—can spot a forged destination before the rack leaves. Ars also points to the related case against Supermicro co-founder Wally Liaw, involving alleged shipments worth $2.5 billion since 2024. Export controls live or die in day-to-day operations. Database Trends and Applications writes:
Under a multi-year $240M agreement between IBM and Together AI, IBM is positioned to deploy a large cluster of NVIDIA HGX B300 systems on IBM Cloud with expected availability in Q1 2027, according to the companies. Together AI will use this cluster to provide open-source model inference.
The IBM-Together deal is $240 million, runs multiple years, and brings HGX B300s online in Q1 2027—refreshingly concrete. But Together AI still has to turn that capacity into token prices low enough to make enterprises reconsider proprietary APIs, not just celebrate a very expensive cluster. And let's be clear about the “open-source” label. Together controls the fine-tuning and inference layer; IBM Cloud owns the physical substrate. Open models don't magically erase control over the stack. NVIDIA says Spectrum-X and B300 deliver 30 times prior-generation AI-factory output. Fine—show me the cost per useful million tokens under enterprise traffic, with real latency and ugly long-context workloads. Q1 2027 is when this gets judged. After the $13.7 billion Georgia contract we just covered, $240 million sounds modest. It isn't. This one names the customer and specifies the hardware. We even get a delivery window—commercial terms are finally getting more legible. This one's from iAfrica:
Nigeria has unveiled a National Digital Cloud Policy aimed at attracting $750 million in private investment within 24 months, covering data centres, cloud infrastructure and artificial intelligence computing capacity, while setting new rules for government cloud adoption and the protection of sensitive public data.
Nigeria wants $750 million in private cloud and AI infrastructure investment over 24 months, but the more useful part is the buying mechanism: pool government demand, then route it through a National Digital Marketplace. That can create a real anchor customer instead of a slide deck full of addressable market. Bosun Tijani's policy draws a more practical sovereignty line than the usual blanket-localization reflex: sensitive government and regulated data stays under national control; commercial workloads can go where the economics work. That's a grown-up distinction. They've also split rule-writing, service delivery, and procurement compliance across three bodies. Good architecture on paper. The announcement still doesn't identify those institutions, so the next test is whether a ministry with a cloud budget can actually buy capacity without the whole thing turning into a committee. It also puts control of infrastructure right on the procurement table. Nigeria can call an inference stack open-source all day; if an outside provider controls the fine-tuning, serving layer, and hardware underneath, sovereignty has limits. This one's from Ars Technica:
Data centers have become the “killer application for solid-state transformers right now,” Lukic told Ars. Several US companies, including Amperesand, Heron Power and DG Matrix, have collectively raised more than $280 million in funding over the past year to commercialize solid-state transformer technology.
The RUM deal gave us the finance side of capacity. Here's the physical choke point: getting power from the grid to a rack without all the usual conversion gear in between. Ars has a real proof point here: Lukic's team put a solid-state transformer on a live grid line and ran a megawatt through it. That's beyond a glossy render. Amperesand, Heron Power, and DG Matrix have raised $280 million around this in a year. Shaving copper, floor space, and electrical complexity is suddenly a very bankable idea. One megawatt is a meaningful hardware test, but a modern AI campus needs this gear manufactured, installed, and maintained at brutal scale. Now they have to prove they can deploy it repeatedly. If you’re finding AI Daily Briefing useful, please subscribe or leave us a review wherever you’re listening. Reviews help other people find the show, and they mean a lot to our team.
We'll be watching the IBM Cloud NVIDIA HGX B300 cluster for Together AI, expected in Q1 2027, along with Nigeria's 24-month window to attract $750 million in private cloud and AI infrastructure investment.
Links to every story are in the show notes. Dig into whichever ones you'd like to explore further. That's today's AI Daily Briefing. This is a Lantern Podcast.