Edge AI Daily Morning Brief: Binance's $6.36 Billion Heist, OpenAI Data Center Renaming, and ChatGPT/Codex Quota Merger

Silicon Valley Frontier:

I. Binance Resumes Withdrawals: Cracks of Trust After a Record $6.36 Billion Heist

1. Incident and Scale: At 20:05 UTC on October 29, 2026, Binance's hot wallet signing system was exploited, and within about 20 minutes roughly $6.36 billion (about $6.3 billion by Binance's own account) was transferred out across 10 blockchains, including about $3.53 billion in Ethereum, about $1.45 billion in Bitcoin, and about $628 million in Tron—all the most liquid mainstream assets—making it the largest theft in crypto history; the FBI, TRM Labs, and Chainalysis all attributed the attack to the North Korean hacker group TraderTraitor.

2. Market and Corporate Reaction: Within 24 hours of the news, total crypto market capitalization fell 3.6% to about $2.68 trillion, Bitcoin dropped to about $77,300 (briefly touching $76,400, a six-week low), BNB fell 6.1%, and more than $1.6 billion in leveraged positions were liquidated; because Binance suspended withdrawals within hours and pledged to bear all losses, the decline was more restrained than during the FTX collapse in 2022. The next day Coinbase shares fell 11% to about $144; its third-quarter revenue was $1.22 billion, down 35% year over year, and of its $718.7 million net profit, about $715 million came from mark-to-market gains on crypto holdings.

3. Recovery, Backstop, and Residual Risk: Tether and Circle froze about $3 billion, but about $2.4 billion (about 420,000 ETH and 18,270 BTC) is not under issuer control and is still held by North Korean actors, and could be absorbed over months through cross-chain bridges and over-the-counter channels; Binance is covering the loss with more than $6.34 billion of its own funds, leaving an actual exposure of about $3.3 billion after deducting frozen assets, while SAFU's roughly $1.23 billion can cover only about one-sixth of the loss—the real protection comes from the company's balance sheet of about $144 billion. The attack escalated from a single-point private key leak to penetration of the hot wallet signing system's permissions, and combined with Coinbase being sold off due to a peer's incident, it exposed the structural fragility of centralized custody and the cracks in user trust.

II. OpenAI Wants to Rename Data Centers "AI Compute Parks"

1. [Background: AI Infrastructure Meets Local Resistance] OpenAI plans to launch a public relations blitz after the November 3 midterm elections to rename its heavily resisted data centers "AI compute parks." According to statistics from the tracking site Data Center Moratoriums on October 12, 27 of the 50 U.S. states are restricting or considering restricting data centers, and only 4 states still offer incentives; New York State in July paused environmental permits for new facilities above 50 megawatts, and bans in Mendocino County, California, and Anne Arundel County, Maryland, run through 2028.

2. [Measures: From "Please Leave Us Alone" to "Let's Set the Rules Together"] OpenAI is rolling out a four-part combination: setting a mandatory "speed limiter" for the most powerful AI systems, signing new benefit contracts with communities, pushing for U.S.-led global AI standards, and renaming them "AI compute parks." In the same period, Amazon and Microsoft successively dropped project nondisclosure agreements (NDAs) as a concession; but critics point out that renaming does nothing substantive about air pollution, noise, water use, and energy costs—the key dispute is who bears the costs and who shares the benefits.

3. [Capital and Flagship Projects] The flagship PORTS-Pike project in Ohio locks in about 8 gigawatts of IT capacity and $4.2 billion in regional grid investment, and during construction through 2032 is expected to create 35,000 construction jobs and 2,500 long-term operating jobs; OpenAI and SB Energy jointly invested $80 million in a community fund and added $84 million in Codex credits for Ohio students. In the same period, Anthropic and OpenAI have injected $100 million into super PACs, but "Leading the Future" sharply cut spending in the final stretch of the midterms, showing that money cannot buy community trust.

III. Apple Plays It Safe on Touchscreen Macs While Reshaping Its Gaming and Software Machine

1. Hardware on the Defensive: On October 28, 2026, Apple released its first touchscreen MacBook Pro, integrating the touch layer into a tandem OLED (PROLED XDR) panel, equipped with M5 Pro and M5 Max chips, starting at $2,899, with preorders opening October 30 and sales on November 6; because macOS has not been rebuilt touch-first, it was reviewed as "stunning hardware, restrained software." Market reaction was muted, with the stock closing down 1.23% at $344.63 and a market cap of about $5.03 trillion, just short of its 52-week high of $351.73, and it was seen as a routine cycle update rather than a value re-rating.

2. Software Forced by Regulators: On August 31, 2026, Apple moved the App Store from Phil Schiller to services chief Eddy Cue, splitting it into two teams—"defend the core" and "adapt"—with Carson Oliver handling daily operations and Ann Thai overseeing third-party distribution, a result of regulatory pressure from the EU, Brazil, Japan, and Australia; new EU rules scrapped the €0.50-per-install core technology fee and replaced it with a 5% core technology commission on digital transactions. Revenue is under pressure: Appfigures data shows U.S. commission revenue has fallen 18% year over year since 2026, and June-quarter services revenue was $30.7 billion, missing the $31.4 billion expectation, with some investment banks already cutting App Store growth forecasts.

3. Games as the offense: At WWDC 2026, Apple launches Game Porting Toolkit 4 and introduces "agentic coding," minimizing the time and cost of porting PC games to Mac, iPad, and iPhone, using in-game purchases/subscriptions to shore up high-margin revenue, drive M-series chips and touchscreen hardware, lock in developers, and rebuild the moat that regulation has weakened. The article suggests investors focus on three metrics: whether the decline in U.S. commission revenue can narrow, whether the share of gaming revenue can rise, and whether the compliance costs of alternative distribution channels are controllable.

IV. OpenAI buys ADT for $15.3 billion: the entry point is more valuable than the model

1. Deal structure: Thrive Holdings, in which OpenAI invests and with which it partners, announced on October 29, 2026, the acquisition of long-established U.S. security company ADT at $10 per share, an equity value of about $7.6 billion and an enterprise value of about $15.3 billion including debt; an Apollo-affiliated fund is providing all-cash financing, with closing expected in the first quarter of 2027. ADT will be taken private and delisted, and will become the cornerstone platform for Thrive's smart-home security business.

2. Target fundamentals: ADT has more than 150 years of history, over 6 million customers, and six 24-hour monitoring centers, but growth has stalled—total revenue in the first half of 2026 was about $2.591 billion, up only 1% year over year, while monitoring services revenue was $2.163 billion, flat year over year; second-quarter recurring monthly revenue was $360 million, customer churn was 13.1%, and the revenue payback period was 2.3 years. What OpenAI values is the physical entry point, trust, and stable cash flow, not growth.

3. Industry and risks: The focus of competition among AI giants is shifting from cloud models to physical-world distribution channels, and the industry is upgrading from "sensors + alarms" to predictive AI. The global smart-home security services market is projected to grow from $37.54 billion in 2025 to $87.29 billion in 2031 (a compound annual growth rate of 15.11%); however, integration difficulty, data privacy (a customer data breach occurred in April 2026), valuation (enterprise value/revenue of about 2.9 times), and competition from giants such as Amazon, Apple, and Google constitute the main uncertainties.

V. Nvidia and Meta prop up the Magnificent Seven, whose earnings growth is overtaken by the other 493 S&P 500 companies for the first time

1. The phenomenon: In the third quarter of 2026, the combined earnings-per-share growth of the "Magnificent Seven" tech companies fell to 22.6%, while the other 493 S&P 500 companies rose to 32.6%, the first time since the fourth quarter of 2022 that the Magnificent Seven's growth was lower than that of the other index constituents; a month earlier (September 30), the two groups' expectations were 33.6% and 27.8%, respectively, a lead of nearly 6 percentage points, and after the two lines crossed in October, the Magnificent Seven fell behind by about 10 percentage points.

2. Structure: The concentration of growth among the Magnificent Seven has risen markedly. FactSet estimates show that excluding Nvidia and Meta, the expected growth of the remaining five would plunge from 33.2% to 7.5%, and the narrative has shrunk from "all seven rising together" to "two carrying five"; meanwhile, the rebound of the other 493 needs to be discounted for quality, as among the top five growth contributors, Micron represents a reversal of the memory cycle, while Meta, EchoStar, and Boeing mainly rely on the low base from one-off charges in the same period last year.

3. Outlook: FactSet forecasts fourth-quarter expected EPS growth of 23.6% for the Magnificent Seven and 30.6% for the other 493, which, if realized, would be the second single-quarter lag since the fourth quarter of 2022; however, as of the end of October, Nvidia and Apple had not yet reported third-quarter results, and the 22.6% is a blend of actual and estimated figures, so their earnings reports could still change the final third-quarter ranking.

VI. SpaceXAI to add invisible watermarks to Grok model text on November 7

1. Background and facts: xAI updated its API documentation on October 31, announcing that starting November 7, seven Grok model IDs (grok-4.7/4.6/4.5/4.3 and the reasoning, non-reasoning, and multi-agent versions of 4.20) will enable invisible text watermarking by default, covering global and regional endpoints and both the Responses and Chat Completions interfaces, citing compliance with Article 50 of the EU Artificial Intelligence Act.

2. Industry comparison: xAI is the last entrant after about 190 signatories including Anthropic (effective August 2) and OpenAI (which launched textGrain on October 5); the difference is that OpenAI enables it by default only within the EU and makes it optional globally via API, whereas xAI extends the default switch worldwide, leaving ChatGPT as the only major North American model not deploying watermarks by default.

3. Mechanism and impact: The watermark forms a statistical fingerprint by slightly adjusting word-selection probabilities when generating tokens, without affecting quality, length, or speed; team administrators can disable it model by model in the console, but must confirm legal exceptions, accept disclosure responsibility, and some models are locked with changes logged in an audit trail. Detection results fall into three categories—"watermarked," "not watermarked," and "uncertain"—with the signal strengthening as text lengthens and decaying markedly after translation or rewriting; xAI has not disclosed accuracy or false-positive rates.

VII. Microsoft XBOX sets a target of exceeding Sony and Nintendo in operating margin by fiscal 2030

1. Goal and starting point: In an internal memo dated October 30, 2026, new Xbox CEO Asha Sharma set a goal of exceeding Sony PlayStation and Nintendo in profit margin by fiscal 2030; yet the Xbox "accountable margin" disclosed in June 2026 was only 3%, versus about 9.9% for Sony's gaming business and nearly 16% for Nintendo, meaning it must rise more than fivefold within four years.

2. Reality and restructuring: In September 2026, U.S. sales of Xbox Series consoles plunged 27% year over year, hitting a record low for September, and revenue fell about 10% in the most recent fiscal quarter; on July 6, 2026, the company launched the "Reset Xbox" restructuring, cutting about 3,200 positions in fiscal 2027 and spinning four studios out of direct control, while Microsoft cut about 4,800 jobs overall (about 2.1% of employees).

3. Strategic pivot: Xbox chose to bypass hardware and bet on monetizing IP, focusing investment on Minecraft and pushing it across film and television, consumer products, and offline experiences (about £50 million invested in the "Minecraft World" theme park at Chessington, UK, in 2027), while also calling out licensing and merchandise opportunities in the Chinese market; according to an October 29 report, the year-over-year revenue decline in the first quarter of fiscal 2027 has already narrowed from 10% to 2%.

1. Apple CEO Ternus set the tone in an October 30, 2026, Wired interview: Apple "doesn't want to launch AI just for the sake of doing AI," arguing that AI should become a foundational capability like chips, screens, and batteries, rather than a product headline.

2. Less than two months into the job, Ternus positioned himself as a "product CEO," continuing Cook's era rather than creating dramatic change; he stressed that product building has a humanistic dimension and that AI agents will not replace product creation.

3. Product moves: At his first launch event on September 9, Apple introduced its first foldable phone, the iPhone Duo (on sale October 23), and on October 13 it released the $299 smart home hub HomeView (shipping November 13); the risk is that if on-device AI and Siri capabilities lag, "AI invisibility" could become a fig leaf for "AI absence."

IX. Nvidia moves the optical engine onto the interposer: before CPO has even settled in, OOI is already knocking at the door

1. On October 13, 2026, at SEMICON West in San Francisco, Nvidia showcased its next-generation co-packaged optics technology OOI: eliminating the organic package substrate and integrating the optical engine directly on top of the silicon interposer that carries the GPU. Samples are in place but it remains at the research stage with no commercialization timeline; this comes less than five months after the May 31, 2026 announcement of volume production for Spectrum-X Ethernet Photonics (CPO).

2. The technical motivation is to shorten the distance between optical-electrical conversion and the compute core: pluggable architectures have electrical loss of up to about 22 dB on 200Gbps channels and per-port power consumption often reaching 30W, while CPO can reduce this to about 4 dB and as low as 9W, a 3.5x efficiency gain; however, CPO's own volume production is still constrained by four factors—manufacturing yield, maintainability, fiber connector standardization, and InP laser supply—and TrendForce expects its capacity ramp-up period to be 2027 to 2028.

3. On the market side, TrendForce (June 2026) forecasts that the CPO/NPO market will expand from about $100 million in 2025 to more than $39 billion by 2030, while pluggable optical modules will still hold nearly $26 billion in 2030, with multiple technologies coexisting long term; Nvidia is using OOI as strategic deterrence, and the real risk for the supply chain is that the segment it is betting on gets restructured and bypassed, with the observation window pointing to around 2027 and the overlap period with the 2028 Feynman GPU.

X. OpenAI merges ChatGPT and Codex quotas into one pool—the good days of subscription pricing are over

1. Timeline and background: On April 2, 2026, OpenAI changed Codex from per-message billing to per-token billing (covering Plus, Pro, and Business), and on April 23 extended it to Enterprise, Education, Healthcare, and Government editions; starting in December 2026, ChatGPT and Codex quotas will be merged into a single token-metered pool.

2. Rules and cost logic: After the merger, quotas are uniformly measured in tokens, and consumption speed depends on the chosen model and task complexity; real-time conversation does not consume quota. API pricing shows GPT-5.5 at $5 input and $30 output (per million tokens), and GPT-5.3-Codex at $1.75 input and $14 output; output unit price is 6 to 8 times the input price, and coding tasks are output-intensive, with the bulk of the bill on the output side. The move is intended to align costs with real compute consumption and dispel the "unlimited" illusion of subscriptions.

3. Impact and commercial intent: The merger hits heavy Codex developers first, because their fast token consumption will also drain their chat quota; the subscription ladder is Go at $8 per month, Plus at $20 per month, Pro at $100 (5x quota) and $200 (20x quota) respectively, and Business at about $20 to $25 per person per month. Once quota runs out, there are three paths: wait for reset, upgrade the subscription, or buy credits; the best hidden option is to switch models by grading the question. The move is essentially laying the groundwork for future "subscription plus usage-based hybrid billing," using quota pooling to shift compute costs while holding user churn down with the promise that "workflows are not disrupted."

XI. Hackers put Claude inside a trojan: Volta automatically profiles and grades victims

1. Cisco Talos researchers found a new information-stealing trojan called Volta (Volta-Autodoxer) in real malware samples. The first round of attacks follows the traditional path: stealing browser passwords, Microsoft OAuth tokens, cookies, payment cards, crypto wallet seed phrases, and active sessions for Discord, Telegram, Steam, and Roblox—the same category as the source of major breaches such as Snowflake and Disney.

2. The trojan's core novelty is "embedded AI": stolen data is sent via OpenRouter to multiple AIs, calling Anthropic's Claude Opus 4.6 by default and cascading down to Moonshot's Kimi K2 and Google's Gemini 2.5 Flash if refused; it also has a built-in 240-word prompt that makes the AI play a "forensic intelligence analyst," automatically generating an executive summary of the victim, assessing their value and operational security, and connecting clues across fields to flag the most valuable and most easily breached targets.

3. Volta has a "sales-oriented" design and may be sold to various attackers, giving less skilled cybercriminals the target screening and intelligence analysis capabilities once reserved for advanced threat actors, using AI to grade credentials in seconds so they can be exploited before tokens expire; CrowdStrike has found four hacker groups using AI to compromise dozens of targets worldwide, but Talos has not yet traced evidence of Volta being used in the wild.