๐ค๐ป AI Daily Digest โ August 7, 2026
Anthropic Puts $10 Billion Behind a Seven-Month-Old Cloud Startup for Vera Rubin Capacity in Norway
Anthropic signed a six-year, $10 billion agreement to buy compute from Volta Infra Holdings, a company founded in January 2026 that raised $300 million at a $2.4 billion valuation, per Bloomberg on August 4. Volta doesn't own the data centers it sells access to. The capacity comes from a 16-year colocation lease with Bitdeer's subsidiary at the Tydal campus in Norway โ roughly 133 megawatts running on NVIDIA Vera Rubin, with two activation phases targeted for the end of this year and March 2027. Bitdeer, the bitcoin miner pivoting into AI infrastructure, disclosed the lease terms in a press release: about $4.7 billion in contracted base revenue, up to $8 billion over 24 years with the optional renewal.
The part I keep circling back to is the financing, not the hardware. Roughly $1.3 billion in standby letters of credit arranged by JPMorgan affiliates backstops Volta's payment obligations to Bitdeer. That's the same credit-decoupling trick Google runs through its TPU lease guarantees: a creditworthy intermediary sits between the tenant and the operator so the operator can borrow cheaper. Volta's founders are ex-Brookfield infrastructure people, and the pitch is basically "compute as a utility" โ you sign for capacity the way you sign for power, and someone else arranges the capital stack.
Critics will call this circular financing โ NVIDIA is both an equity investor in Volta and the chip supplier โ and that concern is fair. It's also true that vendor financing is how aircraft and telecom equipment have always been sold, and that Anthropic is spreading commitments across Google, Amazon, SpaceX, AMD, and reportedly Meta ($10 billion over two years is under discussion) precisely because its single biggest constraint is powered, cooled accelerators. Both things can be true at once: the demand can be real and the structure can concentrate risk. I'd want to see how those letters of credit perform before calling the whole thing either genius or a house of cards.
โ Bloomberg ยท Bitdeer
๐ Bloomberg via Yahoo Finance โ Anthropic signs $10B computing deal with Volta Infra ยท TechTimes โ Anthropic's $10B Norway deal analysis
NVIDIA Opens Alpamayo 2 Super, a 34B Robotaxi Reasoning Model You Can Use Commercially
NVIDIA announced on August 4 that Alpamayo 2 Super โ the second generation of its open reasoning model family for autonomous vehicles โ is now available for commercial use. It's a 34-billion-parameter vision-language-action model: a 32B Cosmos 3 Super Reasoner that interprets up to seven cameras for 360-degree coverage, plus a 2B diffusion-based Action Expert that turns the reasoning into a trajectory. Post-trained with reinforcement learning, it outputs five linked things at once: the planned path, a chain-of-causation trace explaining why, high-level meta-actions (yield, change lanes, stop), visual question answering with 2D grounding, and reasoning auto-labels.
What makes this worth your attention is the license, not the benchmark table. Alpamayo 2 Super ships under OpenMDW-1.1, the Linux Foundation's permissive open model license, covering fine-tuning, derivatives, and commercial redistribution. Earlier Alpamayo versions were research-only; now the whole family is cleared for commercial deployment, and distilled models can be shipped without further permission. The Alpamayo family has passed 500,000 downloads on Hugging Face.
The auto-labeling story is the commercially interesting one. NVIDIA claims the model can take raw fleet footage and generate chain-of-causation labels plus grounded VQA, compressing annotation cycles from months to days. Jensen Huang introduced the model as targeting robotaxis, trucks, delivery vans, and farm tractors. The stated numbers are strong โ 0.911 m minADE on trajectory prediction, 79.2 on LingoQA, 1.50 on the closed-loop AlpaSim score โ but open weights plus a labeling pipeline that runs on your own fleet data is the part that changes how AV companies budget.
โ NVIDIA
๐ NVIDIA Blog โ Alpamayo 2 Super now available ยท NVIDIA Technical Blog โ Generate Trajectories, Reasoning Traces, and Auto-Labels
Liquid AI's 2.6B Model Runs Agents on a Phone and Beats a 9B Model at Tool Use
Liquid AI released LFM2.5-2.6B this week, an open-weight 2.6-billion-parameter model built for on-device agents, with a 128K context window and agentic RL post-training aimed squarely at tool use. The numbers that matter: 220 tokens/s decode on an Apple M5 Max, 113 on a Ryzen AI Max+ 395, and a usable 30 tokens/s on a phone. On a single H100 it sustains nearly 15,000 output tokens per second โ about 1.3 billion tokens a day. In quantized form the model fits under 2.5 GB, which is what makes the phone story real.
The benchmark story is where it gets interesting. LFM2.5-2.6B scores 77.83 on ToolSandbox, ahead of Qwen3.5-9B's 76.44, and 56.88 on BFCLv4, with the smaller Qwen3.5-4B trailing. It tops the instruction-following benchmarks in its class โ IFBench 59.17, Multi-IF 80.07 โ while openly trailing the Qwens on AIME25 and coding. In other words: a tool-execution specialist, not a general reasoning model, and Liquid AI's own benchmark selection says so. It ships with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX, and it runs inside Hermes Agent, OpenClaw, and Pi harnesses.
I keep coming back to what this means for the privacy conversation. An agent loop that plans, calls tools, and executes entirely on a phone never sends your screen content or file contents anywhere. That's a different privacy posture than a cloud agent, not a discount on the same architecture. The honest asterisk is the one Liquid publishes: for coding and open-ended reasoning, reach for something bigger. But "on-device" stopped meaning "toy" a while ago, and this release is a clean data point for that.
โ Liquid AI ยท Hugging Face
๐ Liquid AI Blog โ Deploy Agents Everywhere ยท Hugging Face โ LFM2.5-2.6B
Cloudflare Gives AI Agents an Identity and a Wallet, and the Handle Land Rush Begins
Cloudflare announced on August 4 that it's building the identity and payments layer for agentic commerce: Cloudflare Wallets and cloudflare.pay. The idea is simple and overdue. When an agent shows up at a website to buy an API or sign up for a service, the business on the other end has no reliable way to know who sent it โ existing bot-detection tools were built for search crawlers, not for agents transacting on someone's behalf. Cloudflare's answer is a stable identity: every account gets a unique web address, and you can extend that identity to specific agents, so a receiving business can see who authorized the request.
Then the money part. An account wallet holds stablecoin balances; from it you mint "virtual wallets" for individual agents with guardrails baked in โ a spending cap, an approved merchant list, a maximum transaction size. The payment rail is x402, the open protocol backed by Coinbase, running on USDC micro-payments across Base, Polygon, Arbitrum, World, and Solana. Paired with the Monetization Gateway Cloudflare shipped earlier, this completes a two-sided market: agents buy, businesses sell, and nobody has to manually approve each transaction.
The launch itself turned into a spectacle. Handle registration opened the same day, and X lit up with orange wallet IDs as people grabbed names โ celebrities, brands, AI companies โ with the exact Web3-domain-flipping energy of 2021. CEO Matthew Prince's framing is worth quoting: "When an agent shows up at your door, you need to know who sent it." The guardrails are the part I'll be watching. A wallet an agent can spend from within limits set by a human is useful; the same wallet without enforcement is a fraud surface. Cloudflare is at least starting from the right end.
โ Cloudflare
๐ Cloudflare Blog โ Announcing Cloudflare Wallets ยท Cloudflare Press Release โ Cloudflare Gives AI Agents an Identity and a Wallet
OpenAI's First Consumer Hardware Is a Donut-Shaped Speaker, and Apple Is Trying to Stop It
Bloomberg reported August 6 that OpenAI's first consumer device is a donut-shaped smart speaker roughly the size of a hockey puck, priced between $300 and $400, targeting a 2027 launch. It's designed with Jony Ive's LoveFrom studio โ high-end metal, cameras and environmental sensors, learning the user's habits, with mechanical parts that move on their own to show "life" in the interaction. OpenAI reportedly sees it as the first step toward eventually replacing the smartphone.
The product story is entangled with the Apple lawsuit. Apple claims trade secret misappropriation and has asked for an injunction that could interfere with the launch; OpenAI calls the claims "vague and overbroad," denies any infringement, and filed a motion to dismiss this week. OpenAI says its internal investigation after the suit concluded the product doesn't use Apple's secrets. Apple, meanwhile, says its continuing investigation has turned up 11 more former employees who might be witnesses or involved, beyond the two already named.
I have mixed feelings about this one. A $300-400 ambient AI speaker in 2027 is a real bet that "AI lives in the room with you" is a category people will pay for, and the moving parts are a stab at solving the "where is the thing that feels alive" problem that every stationary speaker fails. But hardware is brutal, OpenAI has shipped nothing physical yet, and the project now carries litigation risk that could slip its timeline. The donut is the first physical test of whether OpenAI can do what Apple did in 2007 โ my honest answer is that nobody outside the company knows yet, and Apple's lawyers are one of the reasons.
โ Bloomberg
๐ ๅๅฐ่ก่ง้ป via NetEase โ OpenAIๆบ่ฝ้ณ็ฎฑๅฎไปท้พ300็พๅ
Sand.ai Open-Sources a 114B MoE Video Model That Activates Just 6B Parameters
Sand.ai released and open-sourced MAGI-2 Preview on August 5, which it bills as the first hundred-billion-parameter open MoE video generation model: 114B total parameters, about 6B activated per token. It sits sixth on Artificial Analysis' image-to-video leaderboard with that 6B active budget, and the cost story is the point โ roughly 0.5 yuan for a 10-second 1080p clip on eight H100s, about a tenth of the per-second cost of mainstream models once the distilled version lands.
Architecturally it's the single-stream bet. Instead of separate audio and video backbones stitched together with cross-attention, text, video, and audio tokens enter one Transformer and exchange information through self-attention at every layer, with shared experts for cross-modal commonalities and modality-specific experts for each stream. That's the same "Speed by Simplicity" single-stream design from Sand.ai's daVinci-MagiHuman paper. The open release includes weights, inference code, and the training system, all Apache 2.0 โ plus MagiAttention and MagiCompiler. The weights run about 307 GB and it wants eight Hopper GPUs, so this is not hobbyist territory, but it's the first time a lab has opened a video MoE of this scale for inspection.
Video generation has been the most API-locked corner of the AI ecosystem โ the frontier stuff ships behind rate limits, priced per token or per frame, and the open community gets the crumbs. A 114B open-weight unified audio-video model breaks some of that asymmetry, even if you need a small GPU cluster to feel it. The efficient-scaling argument is the deeper one: if the MoE path works for video the way it worked for LLMs, the "just add compute" scaling story stops being the only game in town.
โ Sand.ai ยท GitHub
๐ Sand.ai โ MAGI-2 Preview ยท GitHub โ SandAI-org/MAGI-2-preview
OpenAI's Telco Case Study: 65% of Support Tickets Auto-Resolved, ARPU +22%, and the Numbers Come With a Control Group
OpenAI published a customer story on August 3 about Circles, the Singapore company that runs a telco SaaS platform used by carriers across 14 countries on six continents, and also operates its own consumer brand, Circles.Life. The headline numbers: 65% of CareX support interactions are now resolved by AI without a human in the loop, customers who received AI-driven personalization show 22% higher ARPU than the control group, churn is down 9%, and the internal engineering team's development efficiency rose 29% after adopting Codex.
The ARPU and churn figures are worth reading carefully because they compare customers who got AI personalization against customers who didn't โ an actual control group, not a year-over-year comparison. That's rarer in vendor case studies than it should be. The trajectory is the other interesting number: one early carrier hit 55% auto-resolution in the first week, and Circles is aiming at 95% across the full journey as real-time voice comes online. The quote from their global head of growth is worth framing: "AI should empower users, not force-fit them into outdated journeys."
Telco support was supposed to be the category AI couldn't touch โ long-lived accounts, billing edge cases, legacy systems. What this case shows is that the repeatable middle of support, the "check my bill, change my plan, why is my data slow" tier, is exactly where an agentic assistant earns its keep, and that personalization at the point of interaction moves a metric as stubborn as ARPU. The caveat is the selection bias you can't see from the outside โ which customers are offered the AI experience and how the control group is chosen matters as much as the 22%. Still, for anyone building agentic customer service, this is one of the more concrete data points released this year.
โ OpenAI
๐ OpenAI โ Circles powers telco personalization with OpenAI technology ยท dada3c ๆ่งฃๅๆ
Top comments (0)