Editor's note
Today leaned less on flashy new models and more on the plumbing and rules underneath AI and the web. California enacted an independent-audit framework for AI, Nvidia expanded compute in Australia, and DeepSeek advanced an IPO while trimming API prices. On the web side, Chrome shifted to a two-week release cadence and Cloudflare reworked how it treats crawlers — changes that touch how sites are built and who is allowed to crawl them. The through-line: the AI and web stack is being formalized across governance, infrastructure and access.
AIOffice of Governor Gavin Newsom
California enacts first-in-the-nation laws requiring independent AI audits
On September 9, Governor Newsom signed two AI bills, SB 813 and AB 1405. AB 1405 creates a state registry for independent AI auditors and sets standards for their independence, transparency and integrity. The registry must be in place by January 1, 2029, after which only registered auditors may conduct covered AI audits.
Context
The laws aim to prevent AI companies from being solely responsible for evaluating their own technology by building an independent, third-party audit ecosystem. They extend California's recent run of AI and privacy legislation, and define a covered AI audit as an assessment of the internal controls and processes needed to comply with state law. The mandate takes effect in 2029, leaving a lead-in period.
HaLVision's take
Direct impact on SMBs is limited for now — the rules target larger AI systems and the deadline is years out. But the trend toward AI accountability and documentation is likely to grow, so we would advise keeping a simple record of which AI tools you use and where. How the scope of a covered AI audit evolves is worth watching.
WebChrome for Developers
Chrome moves to a two-week release cycle, starting with version 153 on September 8
Chrome 153 shipped as a stable release on September 8, marking Chrome's move from a four-week to a two-week release cycle. The change applies across desktop, Android and iOS. The enterprise Extended Stable channel keeps its existing eight-week cadence.
Context
Shorter intervals mean new features and fixes arrive sooner and in smaller batches. Releases come more often but with less in each one, a strategy meant to reduce the risk of large, disruptive changes. Organizations that value a steadier pace can stay on Extended Stable to slow things down.
HaLVision's take
Faster delivery of platform features — CSS, web APIs — is welcome for builders. The trade-off is more frequent testing, so pre-launch checks need to become even more of a habit for us. For enterprise clients who want to avoid churn, pinning to Extended Stable is the practical recommendation.
WebCloudflare
Cloudflare to split AI crawlers into three types and change defaults from September 15
From September 15, Cloudflare will classify AI crawler traffic into three types — Search, Agent and Training. By default it will block Training and Agent bots on pages that display ads, while allowing Search. It adds a use parameter to Content Signals in robots.txt and applies to new domains and Free-tier customers, with an opt-out window before September 15.
Context
This continues Cloudflare's push to let site owners control AI access and protect the value of their content. What is new is distinguishing crawler intent; preferences declared in robots.txt are reported for compliance through BotBase, and Verified Bots that ignore them or reproduce content in full may lose their verified status.
HaLVision's take
Site owners should decide before September 15 whether to allow or block AI training and agent access. For many SMBs, allowing Search-type crawling — so you can still be found — while restricting Training is a sensible starting default. Our advice is to check your Cloudflare settings first.
AIDIGITIMES
DeepSeek advances a Shanghai STAR Market IPO as it cuts Flash model prices
Reports on September 10 say Chinese AI company DeepSeek is advancing plans for an IPO on Shanghai's STAR Market. It has reportedly engaged several underwriters, including CITIC Securities, with a pre-listing valuation put at around 500 billion yuan (about $75 billion). At the same time it lowered the API prices of its flagship Flash model.
Context
The move is aimed at raising capital for compute, model development and researcher retention, and reflects the broader race for funding across the AI industry. Continued API price cuts also point to an ongoing price war that makes capable models cheaper to use. Final terms of the listing and valuation are not yet settled.
HaLVision's take
For buyers of AI, the practical takeaway is the API price cut seen in DeepSeek's Flash: the cheaper capable models get, the wider the affordable options. We keep price-and-capability comparisons current rather than locking into one vendor, and pick the best combination per project.
AIReuters
Nvidia to build up to 2GW of AI capacity with eight Australian data-center firms
On September 10, Nvidia said it is partnering with eight Australian data-center and cloud companies — Firmus, Sharon AI, IREN, Megaport, ResetData, CDC, NEXTDC and AirTrunk — to build up to two gigawatts of AI factory capacity by 2027. That would more than double Australia's current computing capacity of roughly 1.6 GW.
Context
The move reflects the sovereign AI trend of countries building domestic AI compute, and is anchored to Nvidia's DSX platform, integrating everything from physical infrastructure and compute to networking and software. It is also a sign of sustained demand for data centers and power.
HaLVision's take
Regional compute buildouts mean lower-latency, data-residency-friendly AI closer to home — relevant for clients that need to keep data in-country. We read it as a sign that hosted AI capacity will keep expanding, which over time should ease access and cost for smaller users too.