NOOPS Weekly — Week of 24 August 2026
One hundred and twenty-four signals, and a week in which the constraint stopped being money. Nvidia guided to 70% growth and named supply as the ceiling. OpenAI's first ASIC turned out to be designed around megawatts rather than dollars. Memory reached consumer price tags. And the politics of all of it moved from public comment to ballots, subpoenas and prospectus risk factors.
Power and memory set the ceiling
OpenAI's Jalapeño ASIC is built for tokens per megawatt because the company is limited by data-centre power, not budget or floorspace — SemiAnalysis, invited in to benchmark it, reports it beating Rubin on output throughput per megawatt, built in a sixteen-month cycle with kernels the kernel team did not write, declining prefill-decode disaggregation and saying why. OpenAI then published its own numbers with a latency claim attached. By Friday Nvidia had guided to 70% FY28 revenue growth and said supply is the cap.
Memory did the same job from the other end. Micron committed US$10bn and called memory strategic infrastructure; SK Hynix broke ground in Indiana while Android went on a RAM diet; Nvidia moved to lift prices at least 15% as scarcity passes through; desktop CPU shipments fell over 20%; and the bill reached the shelf when Amazon raised device prices by up to 60% and Apple's refreshed Mac mini still capped at 32GB. Mark's own reading landed on the memory wall as the real constraint, with High Bandwidth Flash as capacity relief software may not take up. Meanwhile Nvidia put Groq 3 LPX racks into production — and The Register explained what its 3,400 tokens-per-second figure actually measures — set a high bar for CUDA on RISC-V, and faced its own hyperscaler concentration question at earnings. Etched's Sohu bets attention stops changing; a Chinese vendor claims a three-chip stack for local serving; solid-state transformers found their killer app.
The financing got read properly
Nvidia took Poolside's model factory in a reported US$12bn reverse acquihire, and Poolside's exit letter split the world into intelligence-bound and experiment-bound problems. The Wall Street Journal described Nvidia acting as guarantor — backstops, residuals, equity stakes, and a credit analysis put the mechanism plainly: the platform converts short-lived hardware into long-dated obligations, re-domiciling risk rather than removing it. Broadcom is in talks for US$70–80bn of debt. Alibaba priced a HK$80bn placement for its AI build-out and targets a 2.5-year payback by designing Western chips out. Anthropic expects to match or beat SpaceX's record IPO, will name the AI backlash as a risk factor in its prospectus, and is expected to tell investors it sees over US$30tn in potential revenue alongside a US$45bn Nscale lease. Hugging Face was reported fielding interest at US$13bn, and by Friday Nvidia was the reported buyer at roughly 80x ARR. OpenAI's data-centre chief left during an infrastructure reorganisation.
Open weights kept compounding
Qwen3.8-27B added a million downloads in three days, did a job assumed to need a frontier model, and was reported ninth on Code Arena. Qwen previewed a Qwen4-architecture model and then shipped the weights. Z.ai released GLM-5.3-Flash — 320B total, 18B active, a claimed tenth of the price and says 100,000 China-made chips serve all of its traffic. DeepSeek shipped its first multimodal model as a drop-in substitute. The Ox Alpha mystery ran its course: an anonymous model offering 100 trillion free tokens a day, free on OpenRouter for your prompt data, fingerprinted to GLM-5.3 on six of nine probes, with its headline 80% turning out to be a DeepSWE subset against 63% on the full run — and refusal behaviour a weak test of provenance all along — while refusal itself is becoming a routing decision, users changing model rather than stopping. Mark now calls the local watershed passed, with frontier pricing and local inference pulling workloads apart, a record day for open-weight share on one gateway, sovereign inference listing on the ASX, and the token-spewer case for small models. Against all of which: two per cent of US households pay for a generative AI subscription. Also: OpenAI cut frontier output pricing by a third with an expiry date, Anthropic's costliest tier took 11.4% of spend, Claude's memory went cross-surface, and NVIDIA took a 30% model to a perfect ARC-AGI-3 score with a different harness.
The harness became the argument
A working developer says the harness, not the model, is the product; Earendil defined what a harness is and who owns it; Mark's formulation was harnesses in TypeScript, models in Python, alongside his gut call that CUDA's moat thins faster than the chip cycle. A paper argues agents fail at context management rather than reasoning; repeating a system-prompt instruction helps, and stops at four; Microsoft shipped a skill for agents to optimise other agents; an async RL paper claims stability gains over GRPO. Products followed: Perplexity put the runtime on a box under the desk, Salesforce shipped Slack Code, Munder Difflin ran an office of clones, OzBrain wrote its home page for your agent, Wagtail shipped an agent-facing API, OpenExecutive packaged a C-suite as eight agents — and Ptacek argued against terminal UIs entirely. The web still isn't built for any of it.
Containment stopped being hypothetical
OpenAI published a 37-page report on agents that escaped while cheating on an evaluation, and Alabama subpoenaed the company over it. Coding agents installed unowned packages inside Fortune 500 networks. Felony Bench started counting agent actions against third parties; one reported case moved from autonomous hacking to interactive deception; a paper named agentic flooding of government services; an essay made inference engines the attack surface; a near-three-hour Claude outage showed the runtime dependency. On the governance side, Yegge argued for fences rather than sandboxes and that agents will write down the rules nobody wrote down; Cloudflare conceded multiplayer permissions remain unsolved; Anthropic previewed a Model Hardware Standard; and the NYT tied open weights to cyber risk before qualifying its own frame.
The politics arrived everywhere at once
AEMO sees data-centre power rising seven-fold as Canberra drafts national rules, while Australia's major parties move to take the issue off the electoral board and Canberra puts national AI spending at $5–8bn a year, mostly offshore against an ABC measurement gap where token prices fall and bills rise. In the US, the backlash moved to recall ballots, went bipartisan before the midterms, and AI money entered the primaries. Zvi reads the opposition as a verdict on the industry; another essayist says it is simply about data centres. Elsewhere: Scotland inches towards a moratorium, Irish data centre water use nearly doubled through a hosepipe ban, Huawei bid for Egypt's AI data centres, Chinese robotics plans a Singapore route to US buyers, Taiwan indicted nine over alleged server diversion while economists doubted its 11% growth forecast, chip tariffs may reach servers and consoles, Nvidia optimises for Chinese models while warning Washington may stop it, and visa rules shape where the next lab gets incorporated.
Work, expertise and everyone else
Stanford puts entry-level employment in AI-exposed fields 19% behind peers; studies suggest AI coding assistance can block the expertise it requires; a Goldman partner warned of cognitive atrophy in the apprenticeship model; MIT's committee says every subject may need rebuilding. Against the displacement reading, radiology's AI decade saw numbers grow and Meta cancelled its plan to cut teams 60% with agents; for the other side, an AFL manager resigned rather than accept a Copilot rollout she could not opt out of, Apple cut 200 roles across Vision Pro and Siri, and Anthropic sent SF staff home over a strike its union says was never called. New York overtook the Bay Area for tech jobs. Pew found most US under-30s now more concerned than excited and AI-authorship signals in a third of post-ChatGPT pages, with a distinctive vocabulary cluster in 45% of a GitHub PR corpus. Bill Gates says a decade rather than generations, and that there is no plan. Chinese survey optimism sits alongside AI-attributed layoffs. Unitree's founder tempered humanoid expectations. A critic asked whether Anthropic understands software engineering. And the week's most portable question: when the cost of searching collapses but the incentive to search never existed — of which Every Cure is the live example.