Published by DogTV ·
Open-source Models Close the Gap — Cost, Performance, and a Shifting Balance
Open-source AI models are matching or approaching closed-source frontier performance at a fraction of the cost, while China's open ecosystem gains momentum.
On July 31, DeepSeek released the official version of V4-Flash through its API documentation. Benchmark results published August 1 by Artificial Analysis give V4-Flash 0731 an Intelligence Index score of 50 — one point below GPT-5.6 Luna's 51, but with a per-task cost 60% lower on DeepSeek's own API. The gap comes from a roughly 98% cache-hit discount, far above the industry-standard 90%. On Arena.ai's Frontend Code Arena, a human-preference benchmark for web development, V4-Flash-High hit 1,586 points, ranking 7th overall and 3rd among open models. Price: $0.14 per million tokens.
On August 4, Alibaba announced Qwen3.8-Max, a 2.4 trillion-parameter model with 95 billion active parameters. Qwen says it will open-source the weights next week — the first time a Max-level model from the Qwen series will be publicly released. API pricing is set at $2 per million input tokens and $6 per million output tokens.
MiniMax released H3, an open-source model that tops video generation benchmarks across both text-to-video and image-to-video tasks. H3 handles text, image, video, and audio in a unified multimodal context, outputting up to 15 seconds of native dual-channel 2K video with audio.
On August 3, Hugging Face CEO Clément Delangue told CNBC that China is "winning the AI race," citing an ecosystem where open models and collaborative development create faster progress than the siloed approach of US frontier labs. He predicted China could dominate both open models and frontier AI by the end of this year or next. Delangue noted that when an unreleased OpenAI agent attacked Hugging Face's platform in July, the company used Zhipu's open-source GLM 5.2 for defense — closed models couldn't be leveraged due to API safety guardrails. Separately, in late July, over twenty companies including Nvidia, Microsoft, Meta, and OpenAI sent a joint letter to the US government urging against restrictions on open-weight AI models.
Infrastructure Spending Hits Record Levels — Hyperscalers, Chips, and Venture Capital
AI-driven cloud revenue is pushing hyperscaler valuations to new highs, while chip startups and venture firms pour billions into the infrastructure buildout.
On August 4, Amazon's market capitalization crossed $3 trillion for the first time, driven by better-than-expected Q2 results. The company reported adjusted EPS of $1.97 against expectations of $1.82, on revenue of $200.6 billion. CEO Andy Jassy raised the 2026 capital expenditure forecast from $200 billion to $220 billion, citing AI infrastructure needs and rising storage chip prices. Jassy told investors the company still lacks enough capacity to meet 2026 demand, and described 2028 demand as "already staggering." Founder Jeff Bezos filed to sell 15 million shares worth approximately $4.1 billion.
On August 3, UK chip startup Olix announced a $312 million Series B at a $3.3 billion valuation — more than triple its $1 billion valuation from February. Founded by 25-year-old James Dacombe, Olix is developing an optical digital processor with a "novel memory and interconnect architecture" for AI inference. The company argues that using general-purpose chips for every stage of token production is reaching efficiency limits, and a shift to specialized chips per stage will unlock a step change in AI cost and performance. Investors include Arm, Hudson River Trading, Netflix co-founder Reed Hastings, and the UK government's Sovereign AI fund. Former Wise CFO Matt Briers joined as CFO; Stanford professor and former Intel executive Nick McKeown joined the board. Olix plans to deliver its first products to customers next year.
On August 1, Index Ventures announced it raised $2 billion in fresh capital across seed, venture, and growth funds, bringing total deployable capital to $3.5 billion. The London-based firm called AI a technology that is "putting the power to build into more hands than ever before" and has already invested in Mistral, Cohere, and former DeepMind researcher David Silver's startup Ineffable Intelligence.
Regulation Arrives, Enterprise Trust Splits
The EU AI Act enters enforcement and Palantir's CEO issues a stark warning about frontier labs — two developments that shape how enterprises adopt AI.
On August 2, the EU AI Act's transparency and risk-classification provisions took effect. Companies must now label AI-generated content and inform users when they are interacting with AI rather than a human. Deepfakes must carry machine-readable markers. The AI Office oversees general-purpose AI model providers, while national regulators handle most other AI systems. Violations can draw fines of up to 3% of annual global turnover. Law firm A&O Shearman partner Peter Van Dyck said the rules are "causing significant concern for US AI labs" and that any lab with European customers is now in scope.
On August 3, Palantir reported Q2 revenue of $1.9 billion, up 93% year-over-year, with $1.1 billion in profit — more profit in a single quarter than total revenue in the same quarter a year earlier. In his shareholder letter, CEO Alex Karp described the AI industry as having "Marxist overtones," arguing that frontier AI labs "intend, knowingly or otherwise, to capture the means of production of their purported partners." On the earnings call, he warned enterprises that paying for AI tokens amounts to funding the migration of their IP and expertise into competing models. Palantir positions itself as model-agnostic, letting organizations control both their data and their AI "exhaust."
Separately, on August 2, reports emerged that Microsoft is testing MAI Realtime, its first natively full-duplex AI voice model. The model supports 16 languages including Chinese, Japanese, and Korean, and offers two voice styles. Unlike turn-based systems, it can listen and speak simultaneously with automatic language detection and mid-conversation switching. No public release date has been set.
On the research frontier, OpenAI disclosed that an internal version of its Astra model made breakthroughs on ten mathematical and theoretical computer science problems that had remained unsolved for at least a decade, with a total token cost of approximately $2,000.