❯ OpenAI Reveals First Benchmarks for In-House Inference Chip Jalapeño — Per-Watt Performance Beats Nvidia GB300

[FIRST BENCH] At the Hot Chips conference on August 25, OpenAI released the first performance data for its in-house inference chip Jalapeño: across three models — GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T — effective compute per watt reaches 1.5x to 1.9x that of Nvidia’s GB200 and GB300 systems, with end-to-end latency 1.7x to 3.6x lower. The chip was co-designed by OpenAI and Broadcom, with tests running on SemiAnalysis’ InferenceX platform.

[HARDWARE CARD] The single package carries 6 HBM4 stacks — 216 GiB of capacity and 15.4 TB/s of bandwidth — rated at 700 watts, yet under measured loads it stays under 550 watts. The rival GB300’s full-card draw is 1,400 watts. Under highly interactive workloads, the gap widens further to 2.1x to 4.1x. A first-gen chip beating Nvidia’s current flagship on this curve is something the past three years haven’t seen.

[CAPACITY KEY] OpenAI’s own deployment cadence is restrained: only a small-batch ramp by the end of 2026, with full scale-out waiting until 2027. So in the near term, Nvidia won’t lose a single unit of shipments — the real change is at the procurement table. When a model company holds an inference chip it can build itself and that’s cheaper per watt, the bargaining chip changes hands. In power-constrained data centers, tokens per watt convert directly into revenue — the only metric OpenAI cares about.

▪ SIGNALA first-gen ASIC beating a current flagship GPU isn’t a process-node win — it’s the ability to put models, compilers, and silicon into the same iteration cycle.

❯ Nvidia notifies customers of ~17% system price hike early next year; Vera Rubin full rack rises to $7.8 million

[PRICE RISE] Server makers have notified customers that Nvidia’s flagship AI systems will rise roughly 17%, covering both Vera Rubin and Grace Blackwell models shipping early next year, per The Information. Morgan Stanley estimates a Vera Rubin VR200 NVL72 full rack at about $7.8 million — nearly double the current GB300 NVL72’s ~$4 million.

[COST BREAKDOWN] The main driver is not the GPU — the report says a single Rubin chip costs only about $50,000. What is really ballooning is memory: per-rack memory cost is now about $2 million, up 435% from the prior generation, roughly a quarter of total rack cost, with LPDDR5X capacity doubled to 54 TB. Long-overlooked passive components are climbing too: printed circuit boards up 233%, chip capacitors up 182%, substrates up 82%. Value across the supply chain is shifting from compute silicon toward memory and components.

[WHO PAYS] On that basis, building a 1-gigawatt data center would cost at least $5 billion more. Cloud providers will most likely pass the increase on to customers renting compute — and the same week, OpenAI just laid out the energy-efficiency math for its in-house chip. Researcher teortaxesTex’s take: on the inference side, Nvidia no longer has a pure technology moat — what remains is its balance sheet and long-term supply-chain contracts. What must be recalculated is the unit inference cost curve for AI application companies: it has been trending down for two years; next year, it may turn up for the first time.

▪ SIGNALThe moment memory eats up a quarter of rack cost, pricing power in AI compute has already begun flowing from Nvidia’s hands toward the memory makers.

❯ Nvidia Groq 3 LPX Racks Enter Full Production, Nebius First AI Cloud to Adopt

[PRODUCTION] Nvidia announced on August 24 that its inference accelerator, the Groq 3 LPX rack, has entered full mass production and will be deployed alongside the Vera CPU and Rubin GPU, going live in the data centers of next-gen AI cloud Nebius within the year. These are the first rack products to actually ship since Nvidia acquired Groq for $20 billion.

[PERFORMANCE] Per the official announcement, a single rack packs 256 LPUs, with 128 GB of on-chip SRAM and 640 TB/s of scale-up bandwidth, built on the MGX architecture with full liquid cooling. In tests by Artificial Analysis using the open-source agentic model Gemma 4 31B, the 100K-token long-context scenario delivers 3,400 tokens/second of output — 4x the closest alternative platform. The previous one-card-handles-all approach is also being split: the Rubin GPU handles context prefill, while the LPU owns the latency-sensitive decode stage, with the two dividing the work inside the same rack.

[AGENT LEDGER] Also in the same announcement: Vera Rubin NVL72 results under real agent workloads. Nvidia says that on SemiAnalysis’ AgentX workload running DeepSeek V4 Pro, throughput per megawatt is up to 30x the previous-gen GB300 NVL72, and per-token cost drops by up to 35x. These numbers are aimed squarely at Jalapeño’s efficiency chart — how many concurrent agents you can sustain under a fixed power budget is replacing peak compute as the No. 1 criterion in data center tenders.

▪ SIGNALBy splitting prefill and decode across two chip types, Nvidia concedes that the era of a single GPU handling the entire inference pipeline is over.

❯ Arm Reveals First In-House Production CPU at Hot Chips: 136 Cores Built for Agent Orchestration

[FIRST IN 36 YEARS] Arm detailed the full specs of its AGI CPU at Hot Chips 2026 — the first time in company history it is selling finished silicon instead of IP. Each chip packs up to 136 Neoverse V3 cores on the Armv9.2 architecture, running at up to 3.7 GHz, built on TSMC 3nm process with a dual-chiplet design. The part exceeds 100 billion transistors with a 300-watt thermal design power.

[SPECS & PARTNERS] Per Arm’s disclosure at the event, each core gets a dedicated 2 MB L2 cache, with an out-of-order execution window of over 384 entries, supporting 10-way dispatch and 8-way commit. Memory and I/O are directly integrated, with latency below 100 nanoseconds. Twelve channels of DDR5-8800 deliver 845 GB/s of bandwidth, with up to 6 TB of memory capacity per chip, plus 96 PCIe Gen6 lanes and CXL 3.0 support. The chip was co-developed with Meta, and OpenAI, Cerebras, and Cloudflare are all on the initial partner list.

[WHY NOW] Arm positions this as the CPU-side orchestrator for agent infrastructure: coordinating accelerators and managing data movement. That role has traditionally defaulted to x86, but agentic workloads have raised its stakes—each task requires repeated tool calls, context switching, and coordination across multiple models, so CPU-side scheduling latency compounds along the inference chain. Intel and AMD’s data center orders now face a competitor that both sets the standard and builds its own silicon. Cloud vendors’ next round of CPU procurement lists will show the first shifts.

▪ SIGNALThe shift from collecting licensing fees to selling chips means Arm is taking business from its own customers. That pivot is worth more than the 136-core number.

❯ Legal AI Firm Harvey Post-Trains Proprietary Model Tenet on Moonshot AI’s Kimi K3

[PIVOT] Legal AI firm Harvey announced its first proprietary model, Harvey Tenet, built on the open-weight Kimi K3 from Moonshot AI, with inference platform Fireworks AI as training partner. Previously, Harvey had always customized on top of closed-source models from OpenAI, Anthropic, and Google; this is the first time it has swapped its base to an open-weight model from a Chinese lab.

[METHOD] Tenet runs on asynchronous reinforcement learning, with training data blending synthetic data, public legal corpora, and human expert annotations. The goal is long-horizon agentic legal work — a single due-diligence engagement breaks down into dozens of steps spanning hundreds of documents. In Harvey’s own published evaluations, Tenet is comparable to frontier models from Anthropic, OpenAI, and Google on multiple legal-agent and knowledge benchmarks. No third-party replication has appeared yet, so these results remain company-reported.

[PROCUREMENT] Harvey closed a $200 million round in March at an $11 billion valuation, co-led by Singapore’s GIC and Sequoia, bringing cumulative funding past $1 billion; its annualized recurring revenue disclosed in January stood at $190 million, with more than 1,300 institutions and 100,000 lawyers using the platform. A US vertical AI company at this scale is now staking its most critical model layer on a Chinese open-weight foundation — the first time a Chinese lab has entered the training stack of a top-tier US applications company.

▪ SIGNALThe decisive factor for open-weight models isn’t the benchmark leaderboard — it’s whether anyone is willing to put their core product on top of them.

❯ WSJ: Anthropic to Pitch IPO Investors on $30 Trillion-Plus Revenue Opportunity

[METHOD] According to The Wall Street Journal, Anthropic will most likely present IPO investors with a $30 trillion-plus potential revenue opportunity, topping SpaceX’s earlier $28.5 trillion. The figure is calculated by converting all human work AI could absorb into annual income, rather than extrapolating upward from software market size.

[PRECEDENT] There is precedent for this kind of narrative: Uber in 2019 pitched $6 trillion at its IPO. NYU finance professor Damodaran had already assessed this kind of algorithm ahead of SpaceX’s IPO, saying it had reached the end of the credible range and was still pushing outward. Anthropic just closed a $30 billion funding round in February at a $380 billion valuation, with 2028 revenue reportedly expected in the $190 billion–$200 billion range and an IPO target valuation of roughly $2 trillion.

[ARITHMETIC] OpenAI, Anthropic, and SpaceX are all telling the same story: AI takes over human work, winner takes all. But these three numbers, each on the order of $30 trillion, overlap one another — no one gets it all. If this truly were a winner-take-all landscape, at least two of the three would see their share approach zero. What cornerstone investors need to ask is not how big this market is, but what entitles Anthropic to a discernible slice of it — the answer will directly determine the IPO pricing range and the pace of subsequent capital expenditure.

▪ SIGNALThe bigger the TAM, the more it indicates the company does not yet have an auditable revenue path to present.

❯ Cook Steps Down as Apple CEO on September 1, Ternus Takes Over and Joins the Board

[HANDOVER] Tim Cook will officially step down as Apple CEO on September 1, with John Ternus, senior vice president of hardware engineering, taking over and joining the board. The arrangement was announced in April and enters its execution phase this week. Cook remains executive chairman, primarily focused on engaging with policymakers in various countries.

[SEND-OFF] On August 23 — the day before the fifteenth anniversary of Cook’s tenure as CEO — Apple held a send-off of about 200 people at the Caffé Macs restaurant on campus, with OneRepublic performing live and Laurene Powell Jobs and Ternus speaking in turn. The Information also reports that Cook’s assistant has already shifted to working with Ternus — the kind of detail that typically means a handover has reached the day-to-day level, not just a change of titles.

[CHALLENGE] In fifteen years, Cook took Apple from $350 billion to the top of global market-cap rankings, built on supply chain and hardware gross margins. What Ternus takes over, though, is a company widely considered a generation behind in generative AI: the Siri overhaul has been delayed repeatedly, an in-house foundation model has yet to appear, and Apple Intelligence’s core capabilities still waver between external partnerships and self-development. The first call to make is build in-house or keep purchasing outside — and that is not something supply-chain management can solve the way swapping a chip can.

▪ SIGNALA hardware engineer takes the CEO seat at a company trailing in the software narrative — Apple is betting its in-house silicon can close the AI gap.

❯ Apple Debuts First 2nm Chip M6, AI Performance Up to 4x M4

[DEBUT] Apple announced the M6 in a press release — its first 2nm chip, powering the new Mac mini. The 12-core CPU has two more cores than the M4 and M5, the 12-core GPU adds two as well, and it offers up to 32 GB of unified memory; Apple says it delivers the “world’s fastest single-threaded performance.” At the same event, the company also debuted a new Mac Studio powered by the M5 Ultra.

[SPECS] The M6 is Apple’s first chip to use three CPU core types at once — 2 ultra cores, 4 performance cores, and 6 efficiency cores — with multi-threaded performance up to 40% faster than the M4. The Neural Accelerator, previously found only in high-end models, now comes to the entry-level Mac mini for the first time; Apple says AI performance reaches 4x the M4, with 2x the graphics performance. The M5 Ultra inside the Mac Studio is Apple’s first quad-die architecture, with up to 36 CPU cores — 12 ultra cores and 24 performance cores.

[LOCAL AI] With 2nm and the Neural Accelerator trickling down, the direct beneficiaries are developers running models locally: an entry-level Mac mini can now shoulder inference loads that previously required the cloud. Perplexity’s on-device agent, released the same day, signaled the same direction. The compute bar for desktops is falling — whether a machine can run mainstream open-source models is the yardstick for personal dev environments in the year ahead.

▪ SIGNALApple didn’t pitch an AI story — it just stuffed the Neural Accelerator into the cheapest machine it sells.

❯ Unitree Sheds 45% in Three Days After Listing, Market Cap Shrinks from US$66 Billion to US$36 Billion

[SHARP RETREAT] Reuters reports that Chinese humanoid-robot maker Unitree Robotics has plunged since its Shanghai listing, with shares retreating about 45% from their peak and market value shrinking from roughly US$66 billion to US$36 billion — about US$30 billion erased in three days. Its August 19 first-day close of 845 yuan was up 460% from the 150.80-yuan IPO price.

[FUNDAMENTAL CHECK] The fundamentals behind this boom-and-bust remain shaky: adjusted first-quarter profit fell about 53% year on year to roughly 40 million yuan on rising costs, a sharp contrast with the company’s full-year 2025 performance. The episode also throws China’s IPO pricing model into relief — new-share supply is constrained on one side, while post-listing intraday price limits are capped on the other, a combination that naturally breeds extreme first-day gains and the ensuing stampede.

[RETAIL ABSORPTION] A 460% first-day gain is not an isolated case — mainland new listings averaged a 225% first-day pop last year. The real problem is the bid beneath: retail investors who chased at the peak are absorbing that US$30 billion gap. Embodied AI has enjoyed smooth primary-market fundraising this year, and Unitree is the first top-tier company to have its valuation publicly priced in the secondary market. Its drawdown will directly compress the listing windows and valuation ranges of those still waiting in line.

▪ SIGNALThe public market priced humanoid robots for the first time — and cut the value in half within three days. The primary market needs a new valuation anchor.

❯ Stability AI Closes $76M Series B as Three Major Labels and Electronic Arts Join

[RARE LINEUP] According to a company announcement, Stability AI, developer of the image-generation model Stable Diffusion, has closed a $76 million Series B round, with the three major record groups — Universal Music, Sony Music, and Warner Music — and gaming company Electronic Arts all entering at once. AMD Ventures and Pacific Alliance Ventures followed on, while existing investors Coatue, Greylock, Sean Parker, and Eric Schmidt added to their positions.

[SUIT TO STAKE] These three record labels were the lead plaintiffs in generative-AI music copyright litigation over the past two years; now they have moved to the shareholder side. The turning point was cooperation before investment: Warner announced in November 2025 that it would build artist-friendly AI tools with Stability, and Universal and Electronic Arts each signed agreements to train models on their own catalogs and intellectual property. Stability subsequently launched Stable Audio 3.0, trained on licensed data, along with a digital audio workstation plugin.

[PRICE OF LICENSED DATA] This round brings Stability’s cumulative fundraising under its current CEO to $232 million. The amount is not large by today’s AI-fundraising standards, but the structure is unusual — content owners exchanged equity for a voice in training-data terms rather than continuing to fight in court. Other model companies still in litigation will take this deal as a benchmark: an equity partnership comes faster than a verdict, and is more controllable than one.

▪ SIGNALRecord labels’ shift from plaintiffs to shareholders shows rights holders have done the math — revenue sharing beats an injunction.

❯ Perplexity Launches Fully Local Portable Computer, Zero Token Cost for Local Steps

[FULL-LOCAL] Perplexity, in partnership with Nvidia, is launching Portable Computer, packaging models, inference engines, agent frameworks, tool sandboxes, and app connectors into a single system that runs entirely on the user’s own hardware. Local steps incur zero token cost. Initial support covers Qwen 3.8 27B and Perplexity’s own post-trained PPLX 27B, with Nvidia Nemotron 3.5 Lightning 30B following shortly.

[HARDWARE] The runtime is Nvidia DGX OS or Ubuntu, with both ARM and x64 supported. RTX GPUs require at least 24 GB of VRAM; the desktop-class Nvidia DGX Spark is the current flagship form factor. When a task exceeds what the local model can handle, it pauses to ask whether to hand off to a cloud large model, rather than silently uploading. The sandbox is hard-isolated at the OS level. Previously, similar local solutions mostly shipped only model weights, leaving the toolchain to be assembled yourself.

[SETUP SAVED] The real value is removing the barrier to “building your own local agent stack”—starting inference services, wiring in tools, configuring a sandbox—which used to take days. Compliance officers managing corporate data egress now have an extra option: sensitive documents can be processed without ever leaving the machine. Subscription AI apps have always billed on cloud invocations; once this local path is viable, the billing metric itself needs a rewrite.

▪ SIGNALZero token cost isn’t about saving money—it turns data staying on-prem from a compliance promise into a physical fact.

❯ Anthropic Unifies Claude Chat and Cowork Memory, Default-On with Manual Opt-Out Only

[MEMORY MERGE] On August 25, Anthropic merged the memory systems of Claude conversations and Claude Cowork: what is learned on one side is immediately available on the other. Memory writes have also switched to real-time updates, no longer waiting for a conversation to end before being saved. The free, Pro, and Max tiers all have it enabled by default.

[SENSITIVE OFF BY DEFAULT] According to official notes, sensitive categories such as health data, race, religious beliefs, political stances, and gender identity are not written to memory by default; users can manually enable the “Include sensitive topics in memory” toggle in settings. Previously, the two products’ memories were independent. Now, to keep them completely separate, there are only three options: use different accounts, pause memory entirely, or use incognito conversations the whole time. The cost of opting out has been raised a notch.

[DEFAULTS DEBATE] The convenience of the integration is direct, but the cost is that the boundary between work scenarios and private conversations disappears. Users who handle company documents in Cowork and chat about personal matters in conversations now share the same memory. The combination of default-on and manual opt-out places the burden of choice on the user side, and enterprise clients’ compliance reviews now need an extra line item.

▪ SIGNALOnce memory is unified across products, users’ calculations of “how much does this company know about me” have to be redone from scratch.

❯ Rumors Say OpenAI Has Finished Pretraining a Trillion-Parameter Foundation Model Bel; No Official Confirmation

[RUMOR] The claim circulating on social media is that OpenAI has completed pretraining for a model codenamed Bel, with total parameters exceeding 10 trillion, a successor to the earlier codename Doug, and potentially the foundation for the GPT-6 generation. The propagation chain is clear: the origin is the account @synthwavedd, amplified by aggregation accounts such as kimmonismus, and OpenAI has made no official statement.

[CREDIBILITY] This is currently a single-source claim, with no frontline journalists following up and no second independent source for cross-verification. The same wave of posts also included claims like “Anthropic’s compute capacity is insufficient to handle OpenAI’s Astra release this year,” which are equally unverifiable. The codename family Bel, Doug, and Astra has never been officially confirmed, and the 10 trillion parameter scale has no corroboration from supply chain or compute procurement.

[WHY IT SPREAD] The rumor has traction because whether trillion-parameter pretraining is still advancing is itself an open question—over the past year, public discussion has shifted focus to post-training and inference-time compute. To truly assess this story, the only evidence that counts is observable side signals such as compute procurement, hiring, and supply chain. Until then, treat it as a signal that the parameter-scale approach has not been abandoned, not as a citable fact.

▪ SIGNALA parameter rumor with no corroboration can dominate the day’s discourse, which itself shows the market is still waiting for a leap at the pretraining level.

❯ OpenAI Restores 5-Hour Codex Limit for ChatGPT Plus; Pro Tier Unaffected for Now

[LIMIT] Starting August 25, OpenAI is restoring the 5-hour usage limit for ChatGPT Plus users on Codex and ChatGPT Work. The limit had been temporarily suspended over the past few weeks, leaving only the weekly cap in place. The $100 and $200 Pro tiers will remain unaffected for the next few months.

[RATIONALE] The engineering leads for Codex and ChatGPT have publicly offered a two-part explanation. First, it smooths compute load, keeping weekly quotas generous. Second, Plus-tier users skew toward newcomers — without an hourly window, they can quietly burn through a week’s quota in two or three days, then end up confused. The weekly cap itself is unchanged; only the 5-hour rolling window has changed. Users still get the same 7-day quota — they just can’t use it all up two days early.

[PRICING] The real price of a large-model subscription isn’t on the price tag — it’s in the quota. The same Plus tier had its limit removed and restored within weeks, showing OpenAI is still using quota parameters to adjust compute supply and demand in real time. Developers buying compute by the hour should treat this as the norm: the pricing page stays the same, but the available quota can change at any moment.

▪ SIGNALSubscriptions are fixed while quotas float — AI services are effectively already priced like spot compute.

❯ Zoom Q2 Revenue Up 4.9% YoY to $1.28B; Q3 Profit Guidance Misses Expectations

[FINANCIALS] Zoom reported Q2 fiscal 2027 revenue of $1.28 billion, up 4.9% year over year, slightly above the $1.27 billion market consensus; enterprise revenue rose 7.8% to $787.5 million, the fastest growth in nearly three years. Adjusted EPS was $1.55, above the $1.48 expected.

[GUIDANCE GAP] The company raised its full-year guidance: revenue of $5.085 billion to $5.095 billion and adjusted EPS of $6.08 to $6.12, both above prior forecasts. But Q3 adjusted EPS guidance missed analyst expectations, with revenue guidance of $1.275 billion to $1.280 billion, roughly flat sequentially — the combination of raising the full year while lowering the current quarter typically corresponds to AI-related spending being concentrated in the second half.

[GROWTH CEILING] The gap between the ~5% overall growth rate and the 7.8% enterprise growth rate is being dragged down by the consumer and SMB business. Zoom has spent the past two years telling the AI Companion story, yet the revenue mix has yet to show signs of being rewritten by AI. Going forward, the metric to watch is whether enterprise growth can hold above 7% for two consecutive quarters — if not, this AI transformation narrative amounts to no more than a tool feature upgrade.

▪ SIGNALThree-year-high enterprise growth can’t move the overall needle; Zoom’s AI story currently holds only among its highest-paying customers.

❯ OnlyFans Paid Creators $6.3B in FY, Net Revenue Up ~10% YoY

[UK FILING] Accounts filed by parent company Fenix International with Companies House show that in the fiscal year ended November 30, 2025, OnlyFans paid out $6.3 billion to creators, with net revenue up ~10% YoY to roughly $1.55 billion and pre-tax profit up 5% to $715 million. Cumulative payouts over the past decade have surpassed $30 billion.

[CREATOR BASE] The filing shows the platform’s creator accounts have reached 5 million. Since 2016, 5,076 people have cumulatively earned over $1 million — roughly 10 in 10,000 of all creators. The payout ratio has held steady at 80/20 (creators keep 80%, the platform keeps 20%), a rarity in the creator economy: the take rate has never been cut as scale expanded. Most content platforms today are raising take rates as they grow.

[BENCHMARK] At a moment when AI content platforms largely can’t explain how creators make money, this filing’s reference value is that it is the only large UGC platform with a fully public revenue-share ledger. $6.3 billion in actual payouts versus $1.55 billion in net revenue — the ratio is out in the open. Any product claiming to remake the creator economy with AI must confront this proven benchmark when presenting its own split terms; the cost of creator migration is baked into that ratio.

▪ SIGNALThe 80/20 split hasn’t changed in a decade — a number that says more about the platform’s bargaining position than any creator-friendly manifesto.