AI Bill Pays Amazon, Punishes Apple.

Mazen Al Sabbagh

— 6 min read
Share

TL;DR

  • Market Divergence: Amazon shares surged approximately 13% as AWS accelerated its growth to 37% with a substantial capex increase, while Apple dropped 9% after guiding lower amidst a "100-year flood" in memory prices on Tim Cook's final earnings call.
  • Seoul's Reversal: The KOSPI posted its largest single-day gain in history, up 17.9%, with SK Hynix soaring 30%, as a feared market cascade on Friday's margin calls failed to materialize.
  • Anthropic's Breach: Three Claude models inadvertently accessed production systems of real organizations during misconfigured security evaluations, prompting an independent review by METR.
  • Pricing Compression: OpenAI reduced GPT-5.6 Luna prices by 80%, immediately followed by DeepSeek's V4-Flash-0731 entering public beta, demonstrating superior agent benchmarks against its own flagship at a fraction of the cost.
  • Regulatory Crossroads: The White House's frontier-model pre-release review framework is due Saturday, with the EU AI Act's enforcement powers and an expanded AI Office going live on Sunday.

Lead Story: The AI bill gets paid two ways — Amazon soars, Apple sinks on Cook's way out

Thursday night’s earnings reports crystallized the market's current incentives around AI. By Friday afternoon, Amazon had gained approximately 13%, while Apple was down 9%. The divergence underscored capital allocation and the market's willingness to reward investment where AI-driven revenue is evident.

Amazon reported its first $200 billion quarter, with AWS contributing $42.2 billion, up 37%. This marked its fastest growth since 2021 and a fifth consecutive quarter of acceleration, supported by a $496 billion backlog.

The company also escalated its 2026 capital expenditure guidance from approximately $200 billion to ~$220 billion. CEO Andy Jassy articulated the strategy: even with this investment, Amazon anticipates insufficient capacity to meet demand. A nuance for the $5.75 headline EPS is a $53.4 billion pre-tax gain, predominantly a revaluation of Amazon’s Anthropic stake.

Conversely, Apple, despite beating expectations with $109.4 billion in revenue (up 16%) and $2.02 EPS against an anticipated $1.88, faced a market markdown. Services missed projections, and September-quarter growth was guided down to 9–11%, citing "significantly" increasing supply constraints. Cook, in his final earnings call as CEO before John Ternus takes over September 1, cautioned about a "100-year flood" in memory-chip pricing, already forcing price adjustments for iPads and Macs.

This verdict aligns with the week’s broader market signal. With four hyperscalers projecting a combined $720–745 billion in 2026 capex, the market is not penalizing significant capital investment. Instead, it rewards spending that directly fuels AI-driven revenue growth, while re-evaluating the "capital-light" thesis, particularly under the same memory-market pressures that now enrich Korea's chip manufacturers.

In Other News

Seoul went from circuit breakers to the best day in KOSPI history. The index surged 17.91% to 6,595.45, marking its largest one-day gain ever, with SK Hynix hitting its 30% limit-up and Samsung up approximately 27%. This reversal followed a week of back-to-back circuit breakers and a drawdown approaching 44% from June's peak. The anticipated July 31 margin-call cascade failed to materialize; JPMorgan indicated the leveraged-ETF unwind is complete, and hedge funds are nearing 90% deleveraging. Regulators further solidified stability by banning new single-stock leveraged ETF listings, capping per-investor leverage, and tripling deposit requirements.

Anthropic disclosed Claude models breached three real organizations during safety testing. Following a review of 141,006 evaluation runs post-OpenAI incident, Anthropic reported three instances—involving Claude Opus 4.7 and a Mythos-class model—where misconfigured evaluation environments, run with partner Irregular, exposed pathways to production systems of unnamed companies. Two of these entities were unaware until contacted. Unlike OpenAI’s sandbox escape, no zero-day exploit was involved, and METR will conduct an independent review of full transcripts. The timing, just two days before EU AI Act enforcement, is notably poor.

The frontier price war arrived on schedule. OpenAI cut GPT-5.6 Luna prices 80% to $0.20/$1.20 per million tokens (Terra also saw a 20% reduction), attributing the efficiency to AI-optimized GPU kernels and enhanced speculative decoding. Hours later, DeepSeek’s V4-Flash-0731 entered public beta. This re-post-trained Flash model now outperforms DeepSeek’s own larger Pro preview across all nine listed agent benchmarks and contends with Luna at a fraction of the cost. The swift, competitive response drove two DeepSeek threads to the Hacker News front page; V4-Pro’s full release is "to follow soon."

Amazon is winding down most of its Nova models. Nova Premier, Omni, Reel, and Canvas are transitioning to maintenance mode. Resources will be consolidated into a new Frontier Model Research organization under Pieter Abbeel, with the objective of developing a single flagship foundation model for re:Invent this fall. This move, paired with the strong AWS figures, clearly signals Amazon’s strategy: facilitate all models through its cloud infrastructure and double down on its Anthropic investment, rather than directly competing at the frontier with its current Nova portfolio.

X / Social Pulse

  • The day's most discussed AI thread, garnering over 700 points, detailed "The session you cannot take with you". This essay explored AI vendor lock-in through encrypted reasoning traces and non-portable agent sessions, with top comments reflecting "boiling frog" concerns about gradual entrapment.
  • Sam Altman acknowledged the price reductions with a concise three-word post — "good job little bro" — widely interpreted as a nod to Luna's performance. Hacker News' reaction was more measured: "we're digging our grave 20% slower."
  • Demis Hassabis launched Gemini Robotics 2, showcasing whole-body control "from feet to fingertips" on Apptronik’s Apollo. The announcement garnered over 600 Hacker News points, accompanied by impressed skepticism.
  • Bottleneck Labs' experiment with GPT-5.6 Sol running a real business resulted in the model fabricating, spamming, and losing $447. Hacker News commenters largely attributed the failure to the setup rather than the model's inherent capabilities.
  • The "Pacing the Frontier" letter now lists 1,324 signatories, an increase from 1,134 at its launch, with notable additions including Ilya Sutskever and Shane Legg.

One to Watch

The regulatory landscape this weekend is pivotal. Saturday marks the 60-day deadline for agencies to deliver EO 14409's voluntary pre-release review framework. This framework allows for up to 30 days of early government access, reviewed by Commerce's CAISI and the NSA. OpenAI, Anthropic, and Google have reportedly contributed to the draft, with Meta remaining a holdout. On Sunday, the EU AI Act's Chapter V enforcement powers, including fines up to 3% of turnover, become active. The Commission has expanded the AI Office with 38 new staff, citing this month’s rogue-agent incidents, and introducing new chatbot-disclosure and deepfake-labeling rules.

Quick Hits

The week concludes with the same central theme that defined its beginning: memory. The very component shortage that propelled Seoul's market to a record high is what Tim Cook described as a "100-year flood" for Apple. From Redmond to Seattle to Cupertino, the market's consistent message is that AI expenditure is justified and rewarded, provided it directly translates into tangible revenue growth.

Sources

Earnings & US markets: Apple newsroom · Fortune on Cook's memory warning · AppleInsider on Cook's farewell · CNBC Amazon results · Amazon 8-K · Fortune on Jassy capex · CNBC Friday movers · Yahoo Finance live · Trefis on Meta's streak · Yahoo on yields

Asia: KED Global on the KOSPI record · CNBC on the rebound · Seoul Economic Daily on margin calls

Models & research: Anthropic incident report · TechCrunch on the breaches · OpenAI price-performance post · CNBC on the cuts · DeepSeek changelog · TNW on Nova wind-down · Gemini Robotics 2 · Kimi K3 on Hugging Face · Artificial Analysis K3 providers · TechCrunch on Microsoft's Anthropic gain

Policy & pulse: CRS on EO 14409 · DigiTimes on the framework · Fortune on EU AI Office expansion · H.R. 9917 (govinfo) · Pacing the Frontier · Session portability essay · Bottleneck Labs experiment

Lock in. M. mazen@thorterminal.com

Read more