8 min read · 14 sources

Every California SaaS invoice gets 7.25% - 10.75% more expensive on January 1, 2027. That is the bill Sacramento just signed into law, and it lands on every vendor with more than $500,000 in California sales. Pricing pages, billing systems, and renewal decks all need an update before year-end.

That tax story sits inside a newsletter that ranges from Google’s AI Mode quietly surfacing pricier products, to Microsoft rebuilding Windows for local AI work, to China handing out AI tokens like mobile data. Here’s what actually moved.

Daily token use in China hit 500 trillion in mid-2026, up from 100 billion in early 2024, and businesses are stapling tokens to coffee, credit cards, and dumplings.

California is taxing your SaaS contract

Source: saastr.com ↗

Governor Newsom signed SB 122 on June 29, and from January 1, 2027 the state extends its sales and use tax to “prewritten” software sold as a service, regardless of whether it ships via download, stream, or browser. The combined rate is 7.25% state plus local district taxes, landing between 7.25% and 10.75% depending on the buyer’s address, and the state is projecting about $2B a year in new revenue.

The carve-outs matter: true custom software, IaaS/PaaS (AWS, GCP, raw compute), human-effort services, and digital media are exempt. But if you sell a SaaS product built once and sold many times, you are inside the tax base. Remote sellers cross into collection duty at $500,000 in California sales, which means most Series A+ vendors already have nexus. Sourcing is to the buyer’s CA address, so two customers in the same product can pay different effective rates, and the vendor collects and remits, not the buyer. Expect price increases of roughly that 8 - 10% range to offset the new cost, plus a sprint to make billing engines destination-aware.

Google AI Mode is a different shopping surface, and a pricier one

Source: productrise.app ↗

Productrise tracked over 2 million product listings across 100,000+ SERPs and AI Mode responses between August 9 and 31, comparing AI Mode against traditional search for the same shopping queries. Products that appeared on both surfaces were 21.6% more expensive in AI Mode on average, with a median price of $149 against $100 in traditional search. Across all listings, AI Mode skewed 49% higher.

The bigger number is overlap: only 1.28% of products in traditional results also showed up in AI Mode for the same query. When prices differed (38.1% of matches), AI Mode was pricier 68.4% of the time, and the main seller differed on 49.6% of matched products. For e-commerce and SEO teams, the implication is that ranking work in classic search says almost nothing about visibility in AI Mode, and the Shopping Graph seems to be surfacing a different (and more expensive) tail. If AI Mode becomes the default answer surface, your pricing and merchandising strategy has to account for two separate indexes.

The incumbents are not going anywhere

Source: x.com ↗

Seema Amble’s thread argues that systems of record (Salesforce, Atlassian, Docusign, Klaviyo) get more valuable as AI agents rise, not less. She cites Salesforce’s “Claudeforce” partnership with Anthropic as proof that general agents are unbundling the interface but still need the underlying data. Her “agent hierarchy” framework separates four agent types by autonomy and judgment, and lands on a sharp conclusion for vertical AI startups: win by performing a specific cross-system job better than an incumbent’s purpose-built agent or a general agent that can reconstruct it.

For engineers building AI-native products, the practical lesson is that competing against an incumbent now requires deeper data assets, specific context, and learning loops tied to one job. A thin wrapper around a general model will get disintermediated the moment the incumbent plugs in its own agent.

Engineering culture is what gets rewarded, not what gets announced

Source: leadership.garden ↗

The piece at leadership.garden pushes back on the standard fix for cultural rot: all-hands speeches, pinned standards docs, and Slack reminders. Munger’s line is the spine: “show me the incentive, I will show you the outcome.” Engineers respond rationally to what gets demoed, what gets celebrated, and what shows up in promotion packets. Product-visible work almost always wins over foundational work like fixing flaky tests, paying down CI debt, or rewriting the build system.

The author’s fix was mundane and concrete: give deep engineering work the same surface area as product wins. Demo slots, all-hands presence, and impact language that names reliability, latency, or test flakiness reduction. Career progression, code quality, and system reliability are downstream of where leadership directs attention, not where leadership directs memos.

Microsoft's Project Zenith puts a 30B model on a developer laptop

Source: blogs.windows.com ↗

Microsoft announced Project Zenith at Build 2026: a Windows 11 experience preconfigured for development on “developer-class devices” with 64 GB+ unified memory and 250+ GB/s memory bandwidth. Windows Terminal and VS Code pin to the taskbar, File Explorer shows extensions and hidden files by default, and long-path support is on. The headline feature is local inference: these machines are specced to run 30B+ parameter models locally and unmetered.

First available on AMD Ryzen AI Halo hardware, with more OEM silicon partners to follow. The strategic angle is that Microsoft is positioning Windows as a viable local-AI development platform, which lets engineers experiment with large models without racking up cloud token bills. It is also a shot at Apple Silicon, which has owned the “run big models locally” story for two years.

GitHub Copilot now orchestrates multiple models under the hood

Source: github.blog ↗

GitHub launched Project HydraFusion as a research preview inside Copilot. It picks one of three execution patterns per request: Single (one model), Cascade (an efficient model drafts, a quality gate escalates if needed), or Critique (one model drafts, an independent critic from a different family reviews, and the drafting model revises). The developer experience is unchanged: select HydraFusion like any other model, and it routes to the right workflow for the task.

In offline evaluations across three agentic coding benchmarks, HydraFusion matched or beat the Opus 5 baseline while reducing estimated workflow cost. The interesting move is that the model-routing complexity is hidden from the user, so cost optimisation becomes an automatic property of the tool rather than something every team has to build themselves. It is available on all Copilot plans during the preview.

China is bundling tokens with coffee, dumplings, and credit cards

Source: restofworld.org ↗

Rest of World reports that daily AI token consumption in China hit 500 trillion in mid-2026, up from 100 billion in early 2024. Chinese open-source and open-weight models cost 60% - 90% less than comparable OpenAI and Anthropic offerings, which is what makes tokens cheap enough to hand out as loyalty rewards. Examples: Moonshot AI’s Kimi credit card with Agricultural Bank of China and Amex in July, China Merchants Bank offering 1.8 billion MiniMax tokens to new AI-developer cardholders in June, China Telecom selling 10 million tokens a month for about $1.40, and a Beijing dumpling restaurant giving out 100+ compute vouchers a day.

The pattern is a consumer-facing distribution and pricing model that does not yet exist at scale in the West: tokens in the slot where airline miles used to live. For anyone pricing AI products, it is a reminder that when marginal cost collapses far enough, tokens become a marketing line item rather than a billable unit.

YC S26 is back to atoms

Source: jaredheyman.medium.com ↗

Jared Heyman breaks down the 245 companies launched so far in YC S26: industrials made up 23% of the batch, against a five-year average of 6.6%. SaaS fell to 12% from a 28% average. B2B software overall is at its smallest share since S22 at 52%. AI-tagged companies are trending down, which may just mean founders stopped labelling themselves as AI since it became the default. The “return of the atoms” framing is real: software margins are compressing, the easy distribution is taken, and the next wave of YC companies is mostly building things that need a factory.

Per-seat pricing is dead, at Basecamp anyway

Source: x.com ↗

Basecamp announced on September 3 that it is ending per-user pricing for new customers, capping plans on projects rather than seats. The new Studio tier starts at $59 a month with unlimited users and agents. Existing customers stay on their current plans. The timing lines up with the Basecamp CLI and an upcoming MCP that let agents act as users; Fried argues charging per agent would be as wrong as charging per person. For tooling buyers, the move removes a friction point for scaling headcount and AI-agent usage without a linear cost increase. Watch whether other team tools follow.

Quick links

Source: threadreaderapp.com ↗

Paul Graham threads on AI displacement: AI displaces a way of working (scutwork) rather than whole occupations, and elite programmers get paid more even as junior roles vanish because they operate “way above” that level. The same thread flags a separate concern about AI-assisted writing eroding the population’s knowledge of how writing is constructed.

Noah Kagan on getting your first three customers: build three lists of people who already trust you, pick the ten most likely buyers, and ask each directly with a price and a deadline.

Dylan of GojiberryAI breaks down channel-by-channel growth by MRR stage: outbound to $6k MRR, Reddit to $25k, LinkedIn and YouTube content to $75k, partnerships and affiliates to $150k, paid ads and hires above that. Channels stack rather than replace.

Growth Unhinged analysed 7,600 AI responses to pricing queries across the Cloud 100 across six engines in late July and early August. Commercial-intent queries on ChatGPT rose from 13.9% to 19.2% year over year, more than two-thirds of software buyers now research via AI engines, and only 57% of companies have AI-readable pricing pages. Your buyers are asking ChatGPT what your product costs, and it is mostly not reading your pricing page to answer.