AI’s cheaper, faster model race accelerates.

The Frontier

Science & tech breakthroughs in AI models, hardware, and computing that define what's possible

Huge data centres rise at “China speed” to power AI ambitions. China is rapidly building data centres in Ulanqab, Inner Mongolia, as part of a broader push to support its AI platforms. The region is shifting from an agricultural centre toward a hub for large-scale computing infrastructure through successive construction phases.

2028 is the benchmark: Ulanqab’s successive data-centre phases position Inner Mongolia as an AI-computing hub beyond its agricultural base. The test is whether this buildout challenges Guizhou’s data-centre dominance by 2028.

BBC

Qualcomm reveals Snapdragon 8 Elite Extreme Gen 6. Qualcomm unveiled the Snapdragon 8 Elite Gen 6 and Snapdragon 8 Elite Extreme Gen 6 mobile platforms. Live Geekbench 6 testing reported scores of 4,293 single-core and 12,924 multi-core for the Extreme version.

Qualcomm has set a high bar for its Extreme tier: 4,293 single-core and 12,924 multi-core. Watch the first Geekbench 6 results after the unveiling for scores at or above both marks; anything lower weakens the Extreme premium.

Mobilesyrup

Kyutai releases Voice of Reason. Kyutai released two open-weight speech-to-speech models built on GLM-4-Voice-9B that solve spoken math without a transcription stage or text model in the loop. The organization said supervised fine-tuning and reinforcement learning raised spoken GSM8K accuracy from 27.3% to 77.1%.

Kyutai’s 49.8-point GSM8K gain—from 27.3% to 77.1% after supervised fine-tuning and reinforcement learning—makes end-to-end voice reasoning credible, not merely novel. The test is whether open-source voice agents adopt the approach within a year.

Marktechpost

Model Developments

Latest AI model releases, capabilities, and technical breakthroughs shaping the competitive landscape

Introducing GPT-6 Sol and Luna. OpenAI introduced GPT-6 Sol and GPT-6 Luna, lower-cost models trained using methods similar to those used for GPT-6 Astra. OpenAI said the models bring advances in professional work, factuality, coding, computer use and alignment to faster, more affordable tiers.

OpenAI’s 50% price cut positions GPT-6 Sol and Luna as a lower-cost challenge to premium AI tiers. The test is whether competitors introduce lower-cost models claiming comparable coding, factuality and computer-use advances by September 2027.

Openai

Anthropic releases Opus 5.5 amid safety debate. Anthropic released Opus 5.5, a cheaper and faster AI model that it said achieved its strongest results to date on its most rigorous internal safety tests. The launch puts model cost, speed and safety evaluation at the centre of its latest flagship-model update.

Anthropic is positioning Opus 5.5 around a three-way claim: lower cost, faster performance and its strongest internal safety result to date. The test is whether its next flagship update can match Opus 5.5 on all three.

New York Times

Capital & Control

Tracking the flow of capital and influence that shapes the tech landscape

Aftermarket driver assist faces federal probe after fatal crashes. The National Highway Traffic Safety Administration’s Office of Defects Investigation has opened a probe into comma.ai aftermarket driver-assistance devices after identifying five crashes involving stopped or slow-moving vehicles. Two crashes resulted in three deaths, while four of the five injured 11 people, some seriously.

comma.ai’s next raise is exposed: ODI’s five-crash probe, including two crashes that killed three people, gives investors a defined diligence trigger. The test before that raise is whether ODI attributes any of the crashes to comma.ai devices.

Ars Technica

TD’s Layer 6 and Cohere begin C$25-million AI collaboration. Toronto-Dominion Bank will invest up to C$25 million over three years in a collaboration between its Layer 6 AI research centre and Canadian large-language-model developer Cohere. The companies will explore secure enterprise AI applications spanning agents, workflow tools, productivity, decision-making and client experience.

TD’s C$25-million, three-year pact funds exploration—not a production rollout—of Cohere and Layer 6 tools for agents, workflows and client experience. The test is whether those efforts yield secure enterprise applications by the term’s end.

Betakit

Infrastructure & Power

The physical backbone of AI: compute, data centers, supply chains, and the geopolitics of energy

OpenAI extends cyber access to Ukraine for civilian defense. OpenAI will provide Ukraine’s Ministry of Digital Transformation and associated cyber-defense teams with access to its Daybreak program to identify software vulnerabilities and develop and test fixes. The program is intended to protect civilian energy, hospital and telecommunications systems amid persistent Russian cyber activity; CERT-UA handled nearly 6,000 incidents in 2025.

OpenAI’s Daybreak access makes Ukraine a measurable test of AI-assisted civilian cyber defense: CERT-UA handled nearly 6,000 incidents in 2025. The test is whether teams identify, develop and test fixes faster; watch for evidence by 2027.

Openai

Digital Defense

Threats and defenses shaping the global security landscape

Volexity spots another China-aligned group exploiting Chrome and Microsoft. Volexity reported that the China-aligned group it tracks as UTA0565 exploited a chain of three zero-day vulnerabilities affecting Chrome and Microsoft products on September 3 and 4, before public disclosure or patches. The campaigns targeted Asian government entities with phishing messages and used spoofed domains impersonating the Center for American Progress and China Digital Times.

Volexity reports UTA0565 used a three-zero-day Chrome/Microsoft chain against Asian government entities on Sept. 3-4, making its zero-day capability the central risk. The test is whether Center for American Progress or China Digital Times spoofed-domain phishing persists before patches land.

CyberScoop

The Performance Edge

Peak performance science for sustained execution in high-stakes environments

Brain scans point to a possible cause of long COVID fatigue. A brain-imaging study led by the Centre for Addiction and Mental Health found evidence linking long COVID to injury in dopamine-releasing neurons. Researchers said the finding may help explain persistent fatigue, reduced motivation, slowed movement and cognitive difficulties reported by some people with long COVID.

CAMH’s imaging evidence puts dopamine-releasing neuron injury ahead of a purely subjective explanation for long COVID fatigue, low motivation and slowed movement. The test is whether further brain imaging reproduces that injury alongside those symptoms.

Science Daily

Market Pulse

Daily market data: indices, sectors, top movers, and volatility

Markets down. The S&P 500 fell 0.5%, the Nasdaq 100 dropped 1.0%, the Dow Jones slipped 0.1%, and the Russell 2000 declined 1.1%, with small caps lagging. Energy led sectors, up 1.6%, while Technology fell 0.9%; Alphabet and Amazon were among the notable decliners. The VIX stood at 16.86, in the normal range.

Alphabet and Amazon’s declines, alongside a 1.1% drop in small caps, point to risk appetite weakening beyond mega-cap tech. Energy’s 1.6% gain is the key offset; the test by next session is whether it remains positive as Technology tries to stabilize.

Yahoo Finance