Ten recent confirmed claims had supporting-article windows of no more than two calendar days, while eight claims from December 2014 remain pending after 4,267 or 4,268 days. Q3 holds 5,207 confirmed claims against 168 denied or failed.
OpenAI committed to 10GW for 20 years while Stripe agreed to pay a reported $7.5B for OpenRouter and Nvidia paired a reported $6B Poolside license with a $1B investment. The week’s biggest bets targeted power, routing, supply, safety, and capital.
Tencent released Hy3’s 295B parameters under Apache 2.0, while Beijing reportedly allowed it to receive about 10,000 Nvidia H200s in mainland China. The contrast shifts competition from owning weights to operating agents at scale.
Harvey reached an $11 billion valuation before launching Tenet, its first in-house legal model. Harvey II places that Kimi K3 adaptation inside matters with files, permissions, task routing, and review.
Reddit says licenses with Google and OpenAI supply about 10% of revenue, while core search reaches more than 70M weekly users. The contest is over who can query the archive, on what terms, and whether people keep replenishing it.
Alibaba’s open models passed 3B downloads in six months, while Hugging Face counted 418M for Google and 227M for Meta in 2026. OpenAI’s enterprise revenue overtook consumer revenue as Nvidia lined up $500B for AI infrastructure and cyber access moved behind vetting.
Amazon backed 7.65 GW of gas generation for one Texas AI campus as SpaceX assigned $15.8 billion of quarterly capex to AI and SK Hynix committed $38 billion to fabs. Safety tests could bind launches, but U.S. oversight stayed private.
Meta’s 20-year agreement for more than 2,600 MW from three Vistra nuclear plants is a more durable demand signal than a headline capacity target. Microsoft’s 20-year, 835-MW Three Mile Island agreement and Google’s 25-year, 615-MW Duane Arnold deal show power procurement moving into AI infrastructure’s commercial core.
Prophesee’s September 2022 Series C brought reported funding to about €130M. In the articles cited here through August 6, 2026, no later report names a customer deployment or production benchmark, while Sony, Qualcomm, Axelera, and Nvidia improved the surrounding edge-vision stack.
Bloomberg reported on August 4 that Volta raised $300M at a $2.4B valuation and said it secured a $10B contract with an unnamed AI developer. The contract is about 33 times the round, but its allocation, refresh, and redeployment rights remain undisclosed.
An OpenAI agent took about 17,600 actions during the Hugging Face intrusion. In the week ending August 2, capability accelerated while containment, six-year capital bets, and verification lagged.
From July 11–13, 2026, three Apple–OpenAI articles produced 12 marks across four confirmed claims; eight claims from December 2014 remained pending after at least 4,246 days.
A reported August move could place ChatGPT and Roblox under the EU’s VLOP regime. The 45 million threshold applies only after a service qualifies as an online platform, leaving ChatGPT’s classification unresolved.
Microsoft committed $2.5 billion to 6,000 field staff as Copilot passed 30 million paid seats and Azure topped $100 billion in fiscal 2026 revenue. OpenAI and AWS are funding parallel deployment units, but their announcements do not trace field work to recurring revenue.
Eleven years after Sony paid $155 million for Toshiba’s sensor unit, Sony and TSMC announced a robot-and-car sensor venture. TSMC’s first Kumamoto fab already makes 12nm-to-28nm chips, but the May announcement does not disclose allocation, yield-learning, or qualification rights.
Apple is reportedly targeting 2026 for a 7-inch hub with facial recognition, a new Apple TV, and a refreshed HomePod mini. Matter turns compatibility into shared infrastructure, putting identity-aware permissions at the center of the smart-home contest.
Nvidia is reportedly discussing a roughly $250B backstop for OpenAI’s lease of a 10 GW Ohio data-center project, while separately planning to invest $1B in Naver’s South Korean site. The chip supplier is using its balance sheet to make future GPU demand financeable.
Q2 logged 5,893 confirmed claims against 1,142 rumors; Q3 through July 26 held almost the same 5:1 ratio. Eight claims from December 2014 remain unresolved after 4,239 or 4,240 days.
Anthropic coverage rose from 114 to 856 articles as enterprise framing reached 19.0% and research framing fell to 22.0%. OpenAI paired classified use with guardrails and Disney IP use with standing oversight.
Apple’s App Store is becoming an AI-abuse enforcement layer as well as a distribution market. San Francisco demanded removal of 13 nudification apps, while Apple reports more than 800 million weekly visitors.
Trump Media is testing political speech as market-data infrastructure. It plans a real-time Truth API after sources say the company pitched fast access to Donald Trump’s posts at up to $100,000 a month.
ASML lifted 2026 guidance to €43B–€45B as Kimi K3 and Inkling widened model access, New York constrained data-center permits, and Europe challenged Android defaults.
Hassabis proposed up to 30 days of prerelease review as Kimi K3 launched with 2.8T parameters and Inkling activated 41B of 975B; distribution, chips, permits and maintainers became the harder limits.
OpenAI burned $3.7B against $5.7B of Q1 revenue while cash rose above $73B. Amazon’s February deal added a $50B investment figure, though its deployment remains unclear.
OpenAI and Anthropic still offer unusually large effective token allowances for $200 a month, even as customers route work to cheaper models and vendors ration supply.
On May 24, DeepSeek made permanent a 75% cut to its flagship model's price. The same corpus carried the other half of the trade: DRAM up nearly 400% since September, NAND up 600%. One boom is deflating the price of intelligence and inflating the price of the matter it runs on. The two numbers are the same number, read from opposite ends.
On the largest US electric grid, wholesale power prices rose 76% year-over-year in Q1 to $136.53/MWh, with data center demand identified as the cause. The same quarter, $80B in AI startup revenue was 89% captured by two companies, and the top 10 US tech billionaires added $550B+ to their combined net worth in 2025. The numerator concentrates. The denominator distributes.
On May 11, OpenAI launched The Deployment Company, a $14B PE-backed joint venture that will compete with Accenture, BCG, Capgemini, and McKinsey on enterprise AI deployment. Seventy-seven days earlier, OpenAI had announced "Frontier Alliances" — a multiyear partnership with the same four firms to do the same thing. Anthropic announced a parallel $1.5B JV with Blackstone, Goldman Sachs, and Hellman & Friedman a week before.
Anthropic planned to grow 10x in 2026. The actual run rate is 80x — an 8x miss against its own raised forecast. The forecast miss forced a $200B Google commitment and a deal to take 100% of SpaceX's Colossus 1, leased from the man Anthropic's founders left OpenAI partly to escape.
In August 2025, the Trump administration converted Intel's CHIPS Act position into a 10% equity stake for $8.9 billion, over Intel's own SEC warnings of "adverse reactions" and Stratechery's published verdict that the deal was "a terrible idea." Eight months later, after a Q1 print drove Intel up 23.6% in a single day — the company's best one-day move since October 1987 — that 10% stake is worth $36 billion. A 304% return in eight months. By any reasonable benchmark, it is the best trade the federal government has placed in any sector in modern memory. The criticism survives the trade. It now has to coexist with the trade.
On April 20, Microsoft disclosed that the week-over-week cost to run GitHub Copilot had nearly doubled since mid-January 2026, paused new Pro and Pro+ signups, removed Anthropic's Opus models from the Pro tier, and began migrating Copilot from flat-rate to token-based billing. The same week, Cursor was reported to be raising $2 billion at a $50 billion pre-money valuation. The company that invented the $10/month AI coding tier just retired it.
On April 17, OpenAI and Anthropic each appeared in nine articles — the first clean tie in a ratio that was 6.5 in 2024. One of OpenAI's nine was the disclosure that a four-month-old startup founded by its ex-engineers just raised $500M from Google Ventures and Nvidia at a $4B valuation. The coverage ratio flipped because the capital ratio already had.
Allbirds sold its shoe business for $39M — down from a $4B+ peak in 2021. It renamed itself NewBird AI, announced an AI compute pivot, and gained 800%+ in a day. The last company to pull this move was Long Island Iced Tea Corp, which became Long Blockchain Corp in 2017 — stock up 500%, then an FBI investigation, then delisting. The question is whether the market rewarding the pivot tells you about the company or about the shortage.
The Ornn Compute Price Index shows GPU rental costs up 48% in two months. Gallup says 50% of US workers have used AI at work. Only 12% use it daily. The infrastructure is priced for a conversion the labor data hasn't delivered.
Five sovereign tech programs from five different countries produced news in a single week. The capital going in is at historic highs. The independence coming out is at historic lows. The ratio — money in, sovereignty out — has collapsed against a denominator nobody in 2015 had to budget against: one year of US hyperscaler capex.
OpenAI's internal documents project $2.4 billion in advertising revenue for 2026, rising to $11 billion in 2027 and $102 billion by 2030 — 36% of total projected revenue. The number that reveals what OpenAI has built itself into.
FBI: US cybercrime losses hit $21 billion in 2025, up 26% in a year and fivefold since 2020. In the same period, every major AI company launched cybersecurity products. Defense spending grew. Losses grew faster. The ratio is the argument.
Average wireless revenue per user was $65 per month in 2000. By 2020 it was under $45 — while data volume per user grew by orders of magnitude. Carriers invested hundreds of billions. The value accrued to Apple, Google, and Netflix. The dumb pipe earns dumb-pipe returns.
AWS S3 storage launched at $0.15 per GB per month in 2006. Today it's $0.023 — an 85% decline across 90+ price cuts. The commodity layer is following the utility cost curve at compressed speed. But AWS margins are above 35%, the managed-services moat is deepening, and AI introduces a market structure the framework hasn't seen.
The competitive ISP market of the 1990s had thousands of providers. A single FCC reclassification in 2005 ended mandatory infrastructure sharing and killed the competitive carriers overnight. The $65 billion federal broadband investment in 2021 was the acknowledgment that private markets had failed — exactly as they failed rural electrification 85 years earlier.
In 1898, Samuel Insull proposed the deal that defined American infrastructure: accept government rate oversight in exchange for a guaranteed monopoly. Allowed returns on equity have fallen from 14% to 10% over four decades. The bargain held. The treadmill didn't stop.
In 1904, over 20,000 independent telephone companies operated in the United States. By 1913, AT&T controlled nearly all of them. The mechanism wasn't better technology. It was network incompatibility — if you can't call your neighbor, the smaller network always loses.
Every American infrastructure — electricity, railroads, telephone, broadband, cellular, cloud — has followed the same five-stage economic cycle. Each completes faster than the last. The telegraph took 22 years to consolidate into monopoly. Cloud took 15.
Microsoft priced Copilot at $30 per seat per month in July 2023. Three years later, its own executive called the number a "major barrier." In the same week, Cursor hit $2B in revenue on usage-based billing, and Anthropic unbundled Claude from the tools that made it useful.
OpenAI burned $15M per day running Sora. The product generated $2.1M in lifetime revenue. That ratio — 7,143:1 — maps the central paradox of AI economics: token prices fall 50x per year, but demand grows faster.
Every article gets a score from five independent signals — position persistence, social authority, coverage breadth, semantic novelty, and gravitational pull. The system borrows from PageRank, information theory, and AUC analysis to answer: how much did this story actually matter?
37%. That's the share of Microsoft's coverage now classified as financial — about money, not products. In 2024, it was 15%. On April 2, Microsoft's AI chief confirmed what the coverage already showed: after $135 billion invested in OpenAI, Microsoft still can't build frontier models. The ratio of capital to capability is the story.
Nvidia averages 2.77 articles per day. On the day Arm announced a competing chip with Nvidia's biggest customers as first buyers, Nvidia appeared in zero. Attention is zero-sum. The flows between entities reveal what headlines can't.