AI compute demand is surging—and suddenly the people building the server racks have receipts. New to this story? Here’s where it stands. Zankore started as a Southeast Asia neocloud plan from Ooredoo, NVIDIA, and Nokia: 200 megawatts of initial AI capacity from Indonesian data centers, built to expand across the region. The shareholders called it a multibillion-dollar commitment and forecast about $13 billion in revenue over five years. The question is how it gets all of that into live capacity. This is AI Daily Briefing. Today, giant cloud commitments and cheaper open-model serving, all happening in a global capacity rush—so who’s taking the risk if the economics shift before the hardware comes online? We'll keep tracking this story — Zankore Southeast Asia AI compute platform. Follow the show so the next update finds you. This one's from 94.7 The Beast:
Aug 11 (Reuters) – CoreWeave lifted its annual capital spending forecast on Tuesday after beating second-quarter estimates, encouraged by a surge in demand for its AI cloud computing services, sending the company’s shares more than 14% higher in extended trading. The AI cloud company also lifted its targets for 2026 revenue and adjusted profit, banking on a ballooning order book, after its revenue backlog topped $100 billion in the June quarter.
CoreWeave just raised 2026 capex to $35 billion to $39 billion, with $104.2 billion in backlog. That capacity build has customers behind it. Customers are attached, sure. But $104.2 billion is future demand priced on today’s serving economics. I’d like the average contract vintage and every customer’s renegotiation language before I throw confetti. They added more than $25 billion in net commitments this quarter and say near-term capacity is sold out. That gives CoreWeave bargaining power with buyers that hyperscalers usually keep for themselves. And revenue more than doubled to $2.58 billion. There’s an actual business here. Still, spending up to $39 billion in one year means the demand curve has to stay very friendly for a very long time. Here's The Economic Times:
The Amsterdam-based AI infrastructure firm reported total revenue of $582.3 million for the quarter ended June, compared with analysts' average estimate of $572.75 million, according to data compiled by LSEG. Revenue at its core AI cloud business, which accounted for about 98% of group revenue, rose more than 500%.
Nebius put up $582.3 million against a $572.75 million estimate, and 98% of that revenue came from AI cloud. That is a very concentrated way to beat expectations. The four contracts averaging more than $1 billion each are the part I’m staring at. Great. But before I celebrate, I want to know when those contracts were signed and how the pricing escalates. And what happens if 2028 inference is radically cheaper than the spreadsheet assumed? Put that beside the CoreWeave numbers, and the signal is clear: neoclouds are landing enormous commitments and raising prices for capacity. The hyperscalers no longer have compute pricing to themselves. Nebius says AI-cloud revenue grew more than 500%, so yes, demand is real. But four billion-dollar-plus deals mean a handful of customers could define the whole forecast. One procurement rethink can hit those numbers fast. This one's from Analytics India Magazine:
IBM and Together AI have signed a multi-year $240 million agreement to deploy a large cluster of NVIDIA HGX B300 systems on IBM Cloud, with the infrastructure expected to be available in the first quarter of 2027. Together AI will use the cluster to provide inference for open-source AI models. The deployment will be the first dedicated large-scale inference cluster on IBM Cloud using NVIDIA HGX B300 systems and NVIDIA Spectrum-X Ethernet networking.
Together AI says it’s already serving 400 trillion tokens a month. This $240 million B300 cluster is a throughput bet with an actual workload behind it. And IBM gets to be the infrastructure answer for open weights at enterprise scale. Together can distribute the models all day, while IBM Cloud owns the B300 cluster and the Spectrum-X network carrying the billable inference. I like the stated goal here: cheaper serving, not another vague promise of bigger models. But Q1 2027 delivery matters—by then, the cost per token customers expect could look very different from the economics in a deal signed this week. CoreWeave and Nebius just put serious demand numbers on the board. Here, we’ve got a named $240 million contract dedicated to inference for 400 trillion monthly tokens. The open-source stack has acquired a landlord. Here's Zhang Weilan at Global Times:
The instance is designed to handle inference for mixture-of-experts models with up to 10 trillion parameters, the company said in a statement sent to the Global Times on Wednesday. This marks the first supernode-form computing architecture in China to successfully run large language models exceeding 2 trillion parameters, according to the company.
Alibaba’s M890 is already on sale in Ulanqab, built for mixture-of-experts inference up to 10 trillion parameters. And it scales the interconnect from 16 cards to 64—because at this size, the networking is the computer. Right—and China’s first supernode architecture to run a model above two trillion parameters marks a real milestone. The leaderboard angle matters less. Alibaba says KimiK3 and Qwen3.8Max are already commercial workloads on it; that’s the detail I care about. IBM and Together are building B300 inference capacity for 400 trillion tokens a month. Alibaba is making the same serving-layer bet on a separate stack in Inner Mongolia. Enterprises may be pricing one AI demand curve against two ecosystems that don’t neatly substitute for each other. Ulanqab’s cold climate helps cool these machines, sure. Commercial availability is the important shift. The hyperscalers don’t get to be the only ones deciding what compute costs anymore. This one's from FutureCIO:
To meet the rising demand for secure, scalable AI computing, the initiative aims to develop one of the region’s largest AI infrastructure platforms, targeting a deployment of 1 gigawatt (GW) of NVIDIA DSX AI Factory capacity over the long term.
Following up on the Zankore compute platform we tracked last week: Indosat now puts the long-term target at one gigawatt. The first 200 megawatts are due in the first half of 2027, with NVIDIA GB300 NVL72 systems. This is more than a regional-tech press release now. I can model 200 megawatts. The one-gigawatt target is ambition. Getting from one to the other means power delivery, cooling, networking, permits, and a demand forecast that has to stay hot for years. The partner split is clear. Ooredoo brings long-term capital, Nokia handles the network layer, NVIDIA brings the stack, and Indosat executes locally. Indonesia is becoming a serious piece of the compute map while a lot of coverage is still staring only at Virginia and Texas. NVIDIA says its DSX MaxLPS can lift computing capacity by up to 40% inside existing power limits. Great, if it holds under real workloads. But efficiency gains don’t erase the grid problem when the destination is a gigawatt. If you want more on AI’s impact in government and defense, check out Anthropic Pentagon Watch, a daily briefing on Anthropic’s fight with the DoD over Claude, military AI use, autonomous weapons, and AI procurement blacklisting—wherever you listen to podcasts.
We’re watching two 2027 milestones. Together AI’s dedicated NVIDIA HGX B300 inference cluster on IBM Cloud is expected in the first quarter. Zankore’s first phase is scheduled for the first half, with about 200 megawatts of GB300 NVL72-powered AI capacity.
Links to every story are in the show notes, so take a look at the ones you’d like to explore further. That’s AI Daily Briefing for today. This is a Lantern Podcast.