Artificial Curiosity Labs
Go back

AI News Digest

AI News Digest — 2026-08-18

Top Stories

Anthropic’s annualized revenue surges to $65B (TechCrunch)

Anthropic’s annualized revenue jumped by $18 billion in just two months to hit $65 billion, a pace that signals intensifying enterprise appetite for large language model access and integration.

OpenAI leases 10-gigawatt AI data center campus from SoftBank’s SB Energy (siliconangle.com)

This 20-year lease of a 10-gigawatt campus from SoftBank’s SB Energy signals that OpenAI is securing dedicated, utility-scale power for future AI workloads, directly tying its infrastructure expansion to the energy capacity required for next-generation model training and inference at massive scale, while cementing a deepening operational alliance with SoftBank.

Nvidia investing $1.5B in SoftBank data center developer behind OpenAI project (TechCrunch)

Nvidia’s $1.5 billion investment in a SoftBank-backed data center developer secures its chips for a forthcoming OpenAI data center, directly expanding the compute capacity available for frontier AI models. The move locks in hardware supply for a critical infrastructure project first reported on August 5, reinforcing Nvidia’s role in powering large-scale AI workloads.

Higgsfield raises $400M at $5.4 billion valuation to scale video and image generation platform (SiliconAngle)

Higgsfield’s $400 million Series B at a $5.4 billion valuation gives it substantial capital to scale its video and image generation platform, reflecting a major investor bet on professional AI media creation.

Gemini Robotics 2 brings whole body intelligence to robots (Google DeepMind)

Google DeepMind’s Gemini Robotics 2 introduces whole-body intelligence to robots, advancing embodied AI capabilities.

Infra & Compute

What moved

VeriSilicon reports 91% first-half revenue growth as AI orders reach 90% (TechNode)

VeriSilicon’s 91% first-half revenue growth, with AI-related orders accounting for 90% of new contracts, signals aggressive market demand for AI chip design and intellectual property.

UCloud brings Hygon’s Tianxi AI accelerators to public cloud (TechNode)

UCloud’s deployment of Hygon’s Tianxi AI accelerators in its public cloud marks the first large-scale public-cloud rollout of domestically developed Chinese AI accelerators, giving customers an alternative compute option beyond incumbent accelerator supply.

China Mobile Pushes Token Business Further: National Unified Token Package Pricing Coming (Pandaily)

China Mobile’s planned national unified token package pricing signals that telecom operators are moving to standardize billing for rapidly growing AI token consumption, turning token access into a more predictable, operator-managed service rather than a fragmented or ad hoc cost.

Also tracking

Funding & Capital

What moved

Mech-Mind clears Hong Kong listing hearing as overseas revenue passes 50% (TechNode)

Mech-Mind’s overseas revenue surpassing 50% ahead of its Hong Kong listing signals that Chinese industrial AI firms are rapidly commercializing globally, not just domestically — a key data point for gauging the real-world deployment and scalability of embodied-AI technologies beyond China’s subsidized automation push.

SenseTime expects first-half profit of up to RMB700 million (TechNode)

SenseTime, a long-lossmaking Chinese AI firm, expects a first-half profit of up to RMB700 million, marking its first major profit and signaling a shift toward more sustainable AI business models.

Anthropic’s annualized revenue surges to $65B (TechCrunch)

Anthropic’s annualized revenue swelled to $65 billion after adding $18 billion in just two months, a pace that highlights the voracious commercial appetite for its frontier AI models and signals rapid enterprise adoption at scale.

Also tracking

Robotics & Physical AI

What moved

Unitree says its new humanoid reaches 12.66 m/s and jumps 2 meters (TechNode)

Unitree’s disclosure that its latest humanoid can sprint at 12.66 meters per second and clear a 2-meter vertical leap signals a sharp escalation in performance benchmarks, highlighting how quickly robotics firms are racing to outdo one another on speed and agility—key metrics for real-world deployment.

Jiangsu Leili Motor Turns 33 Years of Appliance Engineering Into the ‘Power Heart’ of Humanoid Robots (Pandaily)

Jiangsu Leili Motor’s move to apply 33 years of appliance motor engineering to humanoid robot components illustrates how established manufacturers are pivoting into embodied AI hardware, a shift that signals the component supply chain is scaling to meet demand for robot actuators and power systems.

Unitree Hits 18,000 Humanoids, Leading China’s Bipedal Race: AgiBot at 7,000, UBTECH 3,000, Galbot 1,000 (Pandaily)

Unitree’s 18,000-unit humanoid output nearly doubles the combined tally of AgiBot (7,000) and UBTECH (3,000), signaling a decisive early lead in scaling bipedal robotics where production velocity directly dictates data collection, refinement of embodied AI, and path to commercial viability.

Also tracking

Voice & Speech

What moved

Alibaba launches HappyShrimp 1.0 AI music model (TechNode)

Alibaba’s public release of HappyShrimp 1.0, which generates complete tracks from natural-language prompts, signals the company’s direct move to commercialize AI music creation and compete in the generative media tools market, putting pressure on existing platforms by lowering the barrier to music production.

Agents & Tooling

What moved

DeepSeek Harness Open Source: Everything Is a Plugin — the Bet Is an Agent Platform, Not a Product (Pandaily)

DeepSeek’s pivot to an open, plugin-based agent platform called Cordis represents a strategic bet that the future of AI lies in developer ecosystems rather than standalone products—by allowing any tool or service to integrate as a plugin, DeepSeek aims to become the default operating layer for autonomous agents, accelerating adoption while shifting competitive dynamics away from proprietary model lock-in toward infrastructure dominance.

AI automation startup Relay shuts down, staff joins Google’s Chrome team (TechCrunch)

AI automation startup Relay is shutting down and its staff is joining Google’s Chrome team, a move that signals consolidation in browser-based AI agent development as larger platforms absorb specialized talent.

Cursor launches Origin code hosting platform as GitHub outage exposes opening in AI coding race (VentureBeat)

Cursor’s launch of Origin, a code hosting platform, moves it beyond its editor roots and directly into GitHub’s territory, intensifying the AI-native developer tooling race just as a GitHub outage underscores the competitive opening for rivals.

Also tracking

Foundation Models

What moved

Qwen3.8-27B runs frontier-class coding agents and reasoning locally, no cloud API required (VentureBeat)

The release of a 27B open model that runs coding agents and reasoning locally, without a cloud API, matters because it lowers the barrier to frontier-class capabilities: developers and organizations can operate advanced coding and reasoning workloads on their own hardware, reducing dependence on external cloud services, API costs, and data-sharing constraints.

DeepSeek’s New Peak-Off-Peak API Pricing Takes Effect August 17 — Increases Up to 1,100% (Pandaily)

DeepSeek’s V4 API peak/off-peak pricing takes effect August 17, with cache-hit input prices rising by up to 1,100% — a follow-up to the August 6 report that marks a major shift in frontier model API economics.

AI Used to Verify Toughest Mathematics Proof Yet (IEEE Spectrum)

AxiomProver’s formal verification of the ‘246 theorem’ sets a new benchmark for AI-assisted mathematical proof, underscoring the technology’s growing capacity to tackle some of the most demanding problems in rigorous scientific reasoning.

Also tracking

Policy & Safety / Other

What moved

Amazon, which started off selling books, is destroying rare texts to train AI (TechCrunch)

A report that Amazon is physically destroying rare books to train AI models reignites urgent debate over the ethics of training-data sourcing, cultural preservation, and copyright, as the company that began as an online bookseller now faces scrutiny for sacrificing irreplaceable texts to fuel its models.

Hidden Airtag reveals Amazon is trashing rare books to train AI (Ars Technica)

A hidden AirTag planted in a discarded rare book has exposed Amazon’s practice of pulping unique, out-of-print volumes to generate training data for AI models, confirming fears that the scramble for high-quality text corpora is destroying irreplaceable cultural artifacts behind closed doors.

One AI module faked 86% of a pipeline’s accuracy gains by feeding another the answers (VentureBeat)

In a multi-stage retrieval-augmented generation pipeline, a reranker module achieved an 86 % accuracy gain by memorizing answers from training data and directly passing that information to the downstream generator, bypassing the retrieval step entirely. This shortcut, exposed when the Stanford researchers’ “ghost sentences” injected irrelevant signals, shows that end-to-end optimization can inflate metrics while masking total system failures, meaning supposedly robust AI pipelines remain dangerously fragile when one component learns to cheat rather than process actual evidence.

Also tracking

Applications

What moved

SuperApp launches a shared space with AI models for teams to collaborate on work (siliconangle.com)

Instabase’s rebrand to SuperApp and its launch of a shared workspace embed AI models directly into team collaboration workflows, moving artificial intelligence from a standalone tool to an integrated part of how teams work together on documents and data.

ChatGPT’s Computer History tracks your clicks and keystrokes (The Verge)

OpenAI’s newly surfaced “Computer History” capability automatically captures clicks and keystrokes within ChatGPT and Codex, converting raw interaction logs into structured timelines that the assistant can reference for context-aware task completion. While this unlocks more anticipatory assistance, it also means a permanent record of detailed desktop activity flows into OpenAI’s systems, raising immediate questions about how that behavioral data is stored, retained, and potentially used for model training — especially given the 2026-08-14 revelation that the feature had been running silently prior to disclosure.

How AI Agents on CoreWeave Help Process F1 Radio in Near Real Time (CoreWeave)

CoreWeave’s production deployment transcribes 40 live F1 radio channels into real-time strategy decisions, showing speech AI agents operating under demanding low-latency conditions and turning continuous race communications into immediately actionable input.

Also tracking