❯ Jeff Dean’s Discovery Loop in Talks for $1B Raise at ~$10B Valuation

[ASK] Discovery Loop, the new company founded by former Google chief scientist Jeff Dean, is in talks to raise roughly $1 billion at a pre-money valuation of about $10 billion, Business Insider reports, citing people familiar with the matter. The company was only publicly unveiled on August 5, less than ten days after its founding. At that price, it ranks among the highest-valued AI companies ever with no product and no revenue.

[TEAM] Dean spent 27 years at Google. His co-founders are three names of comparable weight: Sanjay Ghemawat (Google senior fellow and Dean’s co-author on MapReduce and Bigtable), Oriol Vinyals (vice president at DeepMind), and Quoc Le (co-founder of Google Brain). The company is registered as a public-benefit corporation, focused on using AI to automate machine-learning research and engineering itself, then extending into hardware design, drug discovery, and clean energy.

[BACKERS] The seed round disclosed at launch was co-led by Radical Ventures and Khosla Ventures, with Kleiner Perkins and Lightspeed participating. Google parent Alphabet is both a founding investor and cloud partner — all four founders decamped, and their old employer paid for the send-off and covered the compute. The $1 billion round has yet to close, and the valuation has not been confirmed by the company.

[PRICE] What $10 billion buys is not a product roadmap — it’s four résumés. If confirmed, the deal re-prices the entire asset class of “top-tier research teams”: the cost of poaching comparable talent rises accordingly, and the market value of a four-person core team gets anchored at the billion-dollar scale. Google, funding its departing people while taking a stake, is signaling it would rather hold this bet outside the company than keep them inside.

▪ SIGNAL A company with no product is worth $10 billion — what’s being priced is the researchers themselves, not the company.

❯ xAI Releases Grok 4.6, Intelligence Index Ties GPT-5.6 Sol, Price Steady at $2

[TIE] xAI released Grok 4.6, scoring 61 on the intelligence index in third-party evaluator Artificial Analysis’s tests, tying OpenAI’s GPT-5.6 Sol and sitting just 1 point behind leader Claude Fable 5. The official pricing page shows pricing identical to the previous Grok 4.5: $2 per million input tokens and $6 per million output tokens.

[GAIN] In that same evaluation, the previously released Grok 4.5 had long been treated as a value option rather than a frontier player in this tier. This round’s progress is concentrated in agentic tasks: DeepSWE rose 11.9 percentage points, Terminal-Bench up 10.3 points, APEX-Agents up 10.4 points, and the composite index gained 5 points. xAI says it leads Sol on CursorBench, FrontierCode, and AA-Briefcase. It is available at launch on Cursor, Grok Build, API, as well as OpenRouter, Vercel, and Cloudflare.

[SPREAD] Typical quotes for same-tier models run about twice that figure — against Fable 5’s $10 input / $50 output, Grok 4.6 scores 98% as high on the intelligence index while coming in about 8x cheaper on the output side. DeepSeek was also cutting prices the same day; together these point to one conclusion: the frontier tier’s price band is collapsing.

[REPRICE] Teams building long-chain agents need to recalculate the per-task cost: approaches previously axed for excessive inference overhead may no longer be ruled out when rerun at the $6 output price. The real squeeze falls on vendors priced a tier up — premiums must be justified by capability gaps, and that gap now sits at just 1 point.

▪ SIGNAL Intelligence gap of 1 point, price gap of 8x — the high-priced tier needs a new justification.

❯ Microsoft’s Maia 300 Debuts This Fall, With Anthropic on the Target List

[TARGET] The Information exclusively reports that Microsoft is betting its next-generation in-house AI chip, Maia 300, can accomplish what previous generations could not — win heavyweight external cloud customers like Anthropic while cutting reliance on Nvidia. The chip is slated to debut this fall, as early as September. Microsoft is in talks with TSMC on capacity, targeting more than 300,000 chips, with delivery in 2027.

[CONTEXT] Microsoft has previously been in early talks with Anthropic about supplying Maia 200 for Claude models, but no deal was reached. The wildcard: Anthropic confirmed this month that it is forming its own chip team — the same company is both a potential Maia 300 customer and a long-term rival on self-developed silicon. Microsoft’s own account: Maia runs internal models and OpenAI models at lower cost, and it is ramping up internal usage through Azure AI Foundry and Copilot.

[THRESHOLD] For cloud vendors, the success of in-house chips has never hinged on tape-out — it hinges on whether they can be sold to external customers. Google’s TPU took nearly a decade to get there; Amazon’s Trainium still gets its volume mostly from Anthropic alone. By writing the hardest-to-crack customer directly into its targets, Microsoft is effectively declaring that internal substitution is no longer enough.

[LEDGER] What this really moves is the buyer-side bargaining structure. A large-model company holding a third viable option negotiates with Nvidia from a different posture — and whether that option holds up will be decided by whether those 300,000 chips are delivered on schedule in 2027.

▪ SIGNAL The watershed for in-house chips is not building them — it is selling them to a buyer who must choose suppliers carefully.

❯ CoreWeave Q2 Revenue $2.58B Doubles YoY, Backlog $104B

[DOUBLE] AI cloud provider CoreWeave posted Q2 revenue of $2.58 billion, up 112% year over year, slightly above the $2.56 billion consensus. Backlog stood at roughly $104 billion as of June 30, up from $99.4 billion at the end of Q1. Shares briefly rose more than 20% after the earnings release.

[POWER] The harder numbers are in power: the company operates 51 data centers with 1.5 gigawatts of total active capacity, adding nearly 500 megawatts in Q2 alone. At that pace, the load it hooks onto the grid each quarter rivals that of a mid-sized city. The company also disclosed that in the first weeks of Q3 it signed over $25 billion in new customer commitments, pushing total backlog to roughly $129 billion.

[COST] The flip side of growth is cash. In the same report, Q2 operating expenses consumed $5.7 billion of free cash flow, and losses kept widening. The $104 billion backlog is a revenue guarantee—but also a construction obligation that must be funded before it can be fulfilled. Every contract maps to racks, power, and chip purchases that have yet to land.

[SPLIT] Market judgment on the company has fully polarized: one camp points to 112% growth and the $129 billion order book; the other is watching its debt costs and customer concentration. The deciding factor is conversion cadence—whether the backlog turns into revenue on contract schedule matters more than how big the backlog is.

▪ SIGNAL The $129 billion backlog is simultaneously an asset and a liability—the only difference is whether it gets energized on schedule.

❯ Former Tongyi Qianwen Lead Lin Junyang Founds Pragmatik Labs at a $2 Billion Angel-Round Valuation

[LEAD] According to multiple media reports, Lin Junyang, the former technical lead of Alibaba’s Tongyi Qianwen, announced on August 12 that he has founded Pragmatik Labs (internally abbreviated as “p7k”) in Shanghai, with a post-money valuation of roughly $2 billion in its angel round and a raise of several hundred million dollars. The company currently has no product and no revenue — barely two months after he left Alibaba.

[LINEUP] Reports say he departed Alibaba this June, with the outside world widely expecting him to strike out on his own. The round was co-led by Gaorong Capital and HongShan, with Tencent and the Shanghai Future Industry Fund participating — a state-backed guidance fund and an internet giant entering the cap table at the same time. Lin Junyang, one of Alibaba’s youngest P10s, steered the core technical line that took Tongyi Qianwen from zero to the world’s most-downloaded open-source model family.

[PIVOT] The direction is the part to watch: the new company will not build foundation models — it will build next-generation agents spanning digital and physical environments. This is a clear route change, redeploying the capital accumulated in pretraining into the next-generation platform bet on agents, and wagering that the model layer’s window of easy gains is narrowing.

[PRICING] The valuation anchor for Chinese AI startups has been raised another notch. The $2 billion is likewise buying a résumé, not a product — structurally the same as Discovery Loop’s pricing logic that same day, just at five times the scale. What loosens next is the hiring market: the retention cost for a big tech firm’s core model lead must now be benchmarked against billion-dollar external options.

▪ SIGNAL A P10 leaves, and two months later commands a $2 billion valuation — China’s big tech talent war just changed its unit of account.

❯ Researchers Use Same-Vendor Weaker Models to Decode Stronger Models’ Encrypted Chain-of-Thought, Netting 182 Credentials

[BREACH] According to Wired, researchers found that feeding a frontier model’s encrypted reasoning traces to a weaker, less-guarded model from the same vendor causes the latter to decode the content into plaintext output verbatim. The process requires no jailbreaking of the strong model, no breaking of the encryption itself, and no access to keys — only standard, unprivileged API permissions.

[STRUCTURE] The problem lies in the design. To protect intellectual property, vendors no longer store chain-of-thought on the server side; instead, it is encrypted and returned to the client, which sends it back in subsequent requests. These encrypted blocks are fully interoperable across sessions, users, and models within the same vendor’s ecosystem — and that interoperability is itself the key. The paper states that all three systems — Claude, GPT, and Gemini — are affected.

[HARVEST] The researchers recovered 315,000 reasoning blocks from public code repositories, extracting 367 pieces of personally identifiable information and 182 credentials — data that any API caller could already have retrieved. In other words, the leak is not a theoretical risk; it is a fait accompli already sitting in public repositories.

[CLEANUP] What needs to be re-audited is the engineering habit of writing reasoning traces into logs or version-control repositories. Developers previously assumed this ciphertext was unreadable and therefore committed and stored it casually; that premise no longer holds. On the vendor side, what needs to change is the scope binding of encrypted blocks — without binding to a session or a model, it amounts to no encryption at all.

▪ SIGNAL The encryption wasn’t broken — what was broken is the default assumption that models from the same vendor can read each other.

❯ Cerebras Q2 Revenue $180M, Up 74%; Full-Year Guidance Raised, Yet Shares Fall 14% After Hours

[GAP] AI chip company Cerebras posted second-quarter revenue of $180.11 million, up 74.3% year over year, but below the market’s $194 million expectation. It simultaneously raised its full-year core revenue guidance to $880 million–$890 million, versus the prior $855 million–$865 million. Guidance raised, revenue missed — shares initially fell more than 14% in after-hours trading.

[STRUCTURE] Breaking it down, growth came almost entirely from the cloud. Cloud and services revenue reached $126 million in Q2, up 281% year over year (287% on a core basis) — a company that made its name on wafer-scale chips now makes its money selling inference services, not hardware. Remaining performance obligations totaled $25.4 billion, which the company called a sign of “extraordinary future demand.”

[PRICING] This is its second earnings report since going public in May. Public-company valuations run on expectations, and the $25.4 billion in performance obligations was already in the price. What the market was really judging that day was current-quarter delivery — the $14 million revenue shortfall offset the $25 million increase in full-year guidance.

[TAKEAWAY] The lesson for fellow compute suppliers is direct: performance obligations are not revenue. Sit on a huge pile of future orders but fall behind on single-quarter delivery, and the market prices you on the latter. What Cerebras needs to fix is its quarterly cadence, not its order book.

▪ SIGNAL $25.4 billion in performance obligations couldn’t absorb a $14 million quarterly miss — public companies are judged quarter by quarter.

❯ Tencent Q2 Revenue Hits 204.8B Yuan, Up 11%, WeChat Ads Drive Growth

[GROWTH] Tencent’s Q2 revenue came in at 204.8 billion yuan (about $30.4 billion), up 11% year over year, slightly above Bloomberg consensus. Net profit was 56 billion yuan (about $8.3 billion), up just 0.7% year over year and down 4% quarter over quarter, below the Bloomberg consensus of 58.4 billion—revenue beat, profit miss is the main theme of this earnings release.

[ENGINE] Advertising is the main growth driver. Marketing services revenue rose 22% to 43.6 billion yuan, which the company explicitly attributes to two factors: AI-driven upgrades to ad recommendation models, and the completion of closed-loop marketing capabilities within the WeChat ecosystem. Gaming spending was also steady. In other words, for now the most direct return on Tencent’s AI investment is showing up in ad recommendation efficiency, not in any single AI product.

[COSTS] The same report answers what ate into profits: capital expenditure surged 190% year over year to 51.8 billion yuan. Excluding AI-related investment, core profit rose 19% year over year—that comparison itself tells the story: the profit slowdown is a deliberate outlay, not a worsening business.

[PACE] The AI ledgers of China’s internet giants are moving from the “investment phase” into the “show returns” phase. Tencent’s answer is to first earn back some money via ad recommendations, then sustain quarterly spending on the order of 50 billion yuan. What investors should watch is the crossover point between capital expenditure and ad growth—when it arrives.

▪ SIGNAL The 22% gain in ad recommendation efficiency is currently the only place where Tencent’s AI spending can be clearly accounted for.

❯ Lovable raises $400M Series C, valuation doubles to $13.3B in seven months

[DOUBLE] Vibe-coding platform Lovable has closed a $400 million Series C at a post-money valuation of $13.3 billion, doubling from $6.6 billion last December in seven months. The round was co-led by Menlo Ventures and EU investment vehicle Scaleup Europe Fund (managed by EQT), with Tencent and other new investors from Asia and Latin America participating.

[REVENUE] The revenue curve backs up the valuation: the company says annual recurring revenue has nearly tripled from $200 million and will approach $600 million by the end of August. The Stockholm-based company only launched in November 2024, and the platform has surpassed 60 million cumulative built projects. The product logic: describe a requirement in natural language and it directly generates runnable software.

[GEOPOLITICS] EU capital co-leading a European AI company is itself signal-bearing — outside the US and China, Europe has for the first time produced a double-digit-billion-dollar sample in the generative application layer, and it did so with industrial-policy capital rather than pure market capital.

[VALUATION] At $600 million in annual recurring revenue, the $13.3 billion valuation implies roughly a 22x price-to-sales ratio — not unreasonable for the AI application layer, provided this nearly-3x growth curve does not turn. What deserves a fresh look is the per-customer revenue assumption for software development tools: when non-developers can also produce deliverable software, the boundary of the paying base shifts with it.

▪ SIGNAL A 22x price-to-sales multiple is not a bet that the tool is good; it’s a bet that the number of people writing software grows by an order of magnitude.

❯ DeepSeek Quietly Lists V4-Pro-0813 with 1M Context and 384K Max Output

[LISTING] DeepSeek has added DeepSeek-V4-Pro-0813 to the “Models & Pricing” page of its official API documentation without a changelog. The flagship version offers a 1M token context window and a maximum output of 384,000 tokens, uses thinking mode by default, and includes a separate non-thinking-mode endpoint for latency-sensitive calls.

[PRICING] According to the API docs, pricing remains at $0.435 per million input tokens, $0.87 for output, with cached input as low as $0.003625. The V4 Flash launched in July had already pulled this tier’s price down, and the flagship now follows the same standard. In the same tier, compared with Grok 4.6’s $2 input, this is more than four times cheaper. However, DeepSeek has warned users that prices may rise significantly later.

[CADENCE] Putting the price list up before any announcement is DeepSeek’s typical release style — letting developers stumble upon it in the API first, then letting community benchmarks complete the narrative. What to watch is that price-increase warning: whether this price is the new normal determines whether teams building long-context applications dare to include it in their cost models for the next year, and whether this price cut is a competitive strategy or a capacity dividend.

▪ SIGNAL The four-times-cheaper flagship model carries its own price-increase warning; cost models need to leave a line for it.

❯ YMTC-Backed Fund Takes Stake in SOI Micro, Betting on Low-Power Silicon-on-Insulator Route

[STAKE] According to the South China Morning Post, the Changcun Industrial Investment Fund — co-founded by Yangtze Memory Technologies (YMTC), China’s largest NAND flash maker — has taken a stake in domestic semiconductor manufacturer SOI Micro. The fund was jointly launched in 2023 by YMTC and the Hubei Integrated Circuit Industry Investment Fund in Wuhan. Its previous investments were concentrated in the memory chip supply chain; this marks its first move into specialty logic manufacturing.

[PERSON] SOI Micro was founded in Guangzhou in 2022 and is led by Ye Tianchun — former director of the Institute of Microelectronics of the Chinese Academy of Sciences, and a past chief technologist for the national “02 Special Project”. The company is developing a low-power logic process aimed at reducing reliance on overseas technology. The silicon-on-insulator route does not chase performance by continuing to shrink linewidths; instead, it engineers the substrate structure. SOI is already commercially mature in RF and IoT chips, but remains a minority approach for general-purpose logic.

[PATH] According to public information, given the restrictions on advanced process equipment, China’s semiconductor industry has only two viable paths: push existing process nodes to their limits, or switch to a technology route less dependent on EUV lithography. Memory-sector capital flowing into logic manufacturing shows the latter path is starting to attract money from within the industry — not just from the national team.

[TIMING] Such bets have validation cycles measured in years and will not change the compute supply landscape in the near term. What deserves attention is the shift in capital flows: when memory makers start putting money behind an alternative logic route, it is a sign that the industry has exhausted its patience with “waiting for equipment restrictions to lift”.

▪ SIGNAL Memory makers’ money flowing into logic manufacturing is a sign that the industry has stopped waiting for equipment restrictions to lift.

❯ Fable 5 Captures 11.4% of Anthropic Revenue in First Month, Token Volume Just 6%

[SHARE] According to U.S. enterprise AI spending data tracked by expense management platform Ramp, Anthropic’s flagship model Fable 5 accounted for 6% of the company’s token consumption and 11.4% of model spend in its launch month — a revenue share nearly twice its usage share.

[PRICING] The gap comes down to price. Under Anthropic’s official pricing, Fable 5 charges $10 per million input tokens and $50 per million output tokens, making it the most expensive tier in the company’s current product line. The same data also shows Anthropic overtook OpenAI in enterprise procurement share for the first time this April. Enterprise customers haven’t rolled out Fable 5 as a general-purpose model — instead, they’re handing it the expensive work: fewer tokens, but every token spent on use cases where clients accept premium pricing.

[CONTRAST] That cuts sharply against the low pricing of Grok 4.6 and DeepSeek V4 Pro announced the same day — the market is moving in two directions at once: the budget tier fights for volume, the premium tier fights for unit value. For model vendors, the key metric on the latter path isn’t call volume but which tasks customers will break precedent to pay top dollar for; for enterprise buyers, the recalculation is which model tier deserves budget allocation.

▪ SIGNAL Usage at 6%, revenue at 11.4% — premium models sell task value, not call counts.

❯ Google Packs Insulin Resistance Trends Into Pixel Watch 5 and Fitbit

[LAUNCH] Google is adding insulin resistance trend monitoring to its Health Guardian feature suite, debuting with the Pixel Watch 5 and rolling out to the Pixel Watch 3, 4, and Fitbit Air this fall. It’s a step by the consumer-electronics company toward non-invasive glucose monitoring, and blood pressure trend tracking is part of the same rollout.

[MECHANICS] Google’s August 12 note shows Health Guardian doesn’t measure blood glucose directly. The watch collects passive sensor data — resting heart rate, heart rate variability, activity volume, skin temperature changes during sleep, and respiratory rate — and feeds it to an AI model on the Fitbit health platform, which estimates metabolic trends over a rolling one-to-six-week window. When it flags a sustained elevation, it pushes an alert; definitive testing is still left to hospital labs. Apple Watch already collects two of those signals — heart rate and heart rate variability — but stops short of metabolic-level inference. Google’s blood pressure trend feature is in the same batch, and both land this fall on the Fitbit Air and Pixel Watch 3 and later.

[LIMITS] Reading metabolic state from indirect physiological markers is an end-run around the hard problem of non-invasive glucose monitoring. Accuracy and regulatory classification are both unresolved, and Google positions this as a trend alert, not a diagnosis. For the wearables industry, though, the competitive battlefield has shifted from step counting and heart rate to early chronic-disease signals. Whoever first makes these inferences credible to doctors takes the next round of pricing power.

▪ SIGNAL Google didn’t solve non-invasive glucose monitoring — it just reframed the question so it didn’t need solving.