The chip trade unwinds again, two rival bets on open-weight scale ship in one week, and Washington starts deciding who gets frontier AI.
AI Tape · indicative weekly close · wk of Jul 21
GOOGL▼ 0.8%
$351.99
NVDA▼ 0.9%
$203.28
AVGO▼ 2.9%
$378.16
MSFT▲ 5.0%
$402.29
AMZN▲ 1.9%
$249.99
META▼ 2.3%
$645.85
COIN▲ 1.7%
$160.43
SPCX▼ 9.8%
$123.99
Bottom line up front
The chip trade unwinds again. The same hardware names that snapped back last week gave it all back and more — Super Micro −14.1%, Micron −11.1%, AMD −9.6% — as reports that SK Hynix is slowing its HBM expansion revived fears that AI capex won't pencil out.
Two rival bets on what beats the frontier, both open-weight, both this week. Moonshot AI's 2.8-trillion-parameter Kimi K3 is now the largest open model ever shipped; Mira Murati's Thinking Machines answered with Inkling, its first model — and admitted upfront it isn't the strongest one available.
Washington starts gatekeeping frontier AI. A draft "Gold Eagle" framework would give federal agencies a 30-day pre-release look at new frontier models — after the White House already blocked, then unblocked, Anthropic's Mythos 5 and gated OpenAI's GPT-5.6 rollout this summer.
01 — The Briefing
Must Know
The three stories you cannot be caught not knowing in a leadership meeting this week.
1
The chip trade unwinds again — and this time the software side of AI shrugged it off
SK SK Hynix Super Micro AMD
Two weeks after AI hardware stocks snapped back from their first-half rout, they rolled over again — harder. Reports that SK Hynix is slowing its high-bandwidth-memory expansion revived the "AI capex won't pencil out" fear that started the whole cycle in June, and profit-taking after a blistering 12 months did the rest. Super Micro fell 14.1% on the week, Micron 11.1%, AMD 9.6%, Intel 9.3%, Arista 7.4%. Fifteen of the 21 AI-relevant names we track closed the week lower. But the divergence is the real story: cloud and enterprise-software names that don't carry hardware capex on their books kept climbing — Microsoft +5.0%, Palantir +5.7%, Salesforce +4.8% — even as chips fell.
Why it mattersThis is the third capex-confidence swing in six weeks. If your budget or vendor roadmap is reacting to any single week's tape, you're steering by noise — the split between hardware and software names this week is the more durable signal than either week's direction alone.
2
Two open-weight bets on what beats the frontier — maximum scale vs. maximum customization
KIMI Moonshot AITM Thinking Machines
China's Moonshot AI shipped Kimi K3 on Jul 16 — a 2.8-trillion-parameter mixture-of-experts model, the largest open-weight system ever released, with a 1M-token context window and full weights due Jul 27. It topped Moonshot's own benchmark suite against Claude Opus 4.8 and GPT-5.6, and led Arena's blind Frontend Code evaluation ahead of Claude Fable 5. A day earlier, Mira Murati's Thinking Machines released Inkling, its first model (975B total / ~41B active parameters, trained on 45T multimodal tokens) — and said outright in its launch post that it is "not the most performant model available today, closed or open." The bet: enterprises will pay for a model they can make their own on Thinking Machines' Tinker platform, not the one on top of a leaderboard.
Why it mattersTwo credible labs just staked opposite claims about what wins the next phase — raw scale-at-any-cost versus fine-tunable ownership. Pilot both postures before committing a platform bet to either one.
3
Washington moves from watching frontier AI to deciding who gets it
Anthropic OpenAI Google
A June executive order and a run of case-by-case requests have quietly put the federal government in the loop on which organizations get access to the most capable AI systems — not just the labs. The White House is finalizing a voluntary "Gold Eagle" framework with OpenAI, Anthropic, and Google that would hand federal agencies a 30-day pre-release window on frontier models. It's already been tested: the administration blocked Anthropic's Claude Mythos 5 and Fable 5 in June over national-security concerns before either hit the market — lifting the block June 30 — and asked OpenAI to gate part of its GPT-5.6 rollout to a government-vetted partner list under a program called Daybreak, echoing Anthropic's own curated-partner effort, Project Glasswing.
Why it mattersModel access is no longer just a vendor decision — it's becoming a compliance one. If your roadmap depends on a frontier model your lab hasn't cleared for general release, build in a contingency; the gate can now close before launch day, not after.
02 — Follow the Money
Market Signal
How the week's AI news showed up in the tape — the hardware rebound gave way to a deeper pullback, while enterprise-software names kept climbing.
Story of the week · The hardware pullback deepens
SMCI
$23.83
▼ 14.1% on the week · the hardest-hit AI name as capex-sustainability fears returned
Indicative weekly close · wk of Jul 21, 2026 · not investment advice
This week's movers · hardware gives it back, software keeps climbing
SMCI▼ 14.1%
MUMU▼ 11.1%
AMD▼ 9.6%
SPCX▼ 9.8%
PLTR▲ 5.7%
Memory and server names led the decline — Super Micro −14.1%, Micron −11.1%, AMD −9.6% — as reports of SK Hynix slowing HBM expansion revived capex-sustainability fears. SpaceX kept unwinding its post-IPO pop (−9.8%). Software held up: Microsoft +5.0%, Palantir +5.7%, Salesforce +4.8%.
03 — By the Numbers
The Week in Figures
Six numbers that frame the hardware pullback, the open-weight scale race, and where the capital's still landing.
0.8T
parameters in Moonshot's Kimi K3 — the largest open-weight model ever shipped (Moonshot AI)
0%
weekly decline for Super Micro — the hardest-hit name in this week's AI-hardware pullback (indicative weekly close)
$49B
final close of Abu Dhabi's MGX Fund I, one of the largest AI-focused raises to date
$325M
paid by Hyundai for SoftBank's remaining Boston Dynamics stake, taking full ownership (Bloomberg)
0%
of buy-side finance firms now run generative AI org-wide, up sharply this year (industry survey)
$700M
raised by Neko Health to expand its AI full-body scanning platform into the U.S.
This week's AI-hardware pullback (% decline)
Arista (ANET)−7.4%
Intel (INTC)−9.3%
AMD−9.6%
Micron (MU)−11.1%
Super Micro (SMCI)−14.1%
Breadth of this week's pullback
15 of 21 names fell
71% of the 21 AI-relevant stocks we track closed lower this week (indicative weekly close) — concentrated in chips and infrastructure. The six gainers were all cloud, software, or fintech names with no hardware capex on their books.
04 — On the Wire
Stay Up to Date
The steady developments shaping the field — the open-weight race, the money, and the policy shift.
ModelsKIMI
Moonshot AI ships Kimi K3, the largest open-weight model ever
A 2.8-trillion-parameter mixture-of-experts model with a 1M-token context window, live via API and apps Jul 16; full weights land Jul 27 under an MIT license. China's frontier labs are now setting the pace on open-weight scale, not just catching up on it.
ModelsTM
Mira Murati's Thinking Machines debuts Inkling — and admits it's not the best
975B total / ~41B active parameters, trained on 45T multimodal tokens, open-weighted and live for fine-tuning on the company's Tinker platform. Betting enterprises want a model they can shape, not just rent from a frontier lab.
Markets
The AI-hardware selloff deepens on capex doubts
Reports that SK Hynix is slowing HBM expansion reignited the "AI capex won't pencil out" fear from June; Super Micro, Micron, AMD, and Intel each fell more than 9% on the week. The compute-glut debate isn't resolved — it's recurring on a roughly monthly cycle now.
Policy
The White House drafts "Gold Eagle," a pre-release review for frontier models
Reported by CNBC: the framework would give federal agencies a 30-day look at new frontier releases, building on programs already run inside OpenAI (Daybreak) and Anthropic (Project Glasswing) for gated, government-vetted rollouts. Regulatory review is moving earlier in the release pipeline, not just after launch.
Robotics / M&ABD
Hyundai takes full ownership of Boston Dynamics
Hyundai is buying out SoftBank's remaining ~10% stake for $325M after SoftBank exercised a 2021 put option, valuing Boston Dynamics at $3.3B; Hyundai targets a first Atlas factory deployment in Georgia by 2028. A car maker now owns its humanoid-robotics supplier outright, not just a majority stake.
FundingMGX
MGX closes a $49B AI fund; Together AI raises $800M
Abu Dhabi's MGX finalized Fund I at $49B, one of the largest AI-focused raises to date; Together AI closed an $800M Series C at an $8.3B valuation for its GPU-cloud and model-serving platform. Sovereign and infrastructure capital keeps flowing even as public hardware stocks sell off.
HealthCMS
CMS stands up a new Office of Health Technology Products
The office will oversee AI, interoperability, and other emerging digital-health tools inside Medicare and Medicaid, as UnitedHealth (~$1B) and HCA (~$400M) project AI-driven savings this year. A dedicated federal office for health AI means procurement and compliance rules are coming faster than in most other sectors.
⚑ On your radar — agentic commerce is about to get its first real regulatory scrutiny
France's competition authority spent from January building and deploying its own AI shopping agents, running them through 550 real prompts, then published a 3,700-page advisory opinion on the AI-agent market. Action: if you're deploying customer-facing shopping or booking agents, assume EU-style scrutiny of steering, disclosure, and fairness is coming to that category next — not just to model training data.
05 — The Long Read
Deep Dive
One idea, unpacked — for the leader who wants the "so what," not just the headline.
Strategy · The confidence swing
The AI trade can't decide what it believes — and the capital behind it isn't waiting for an answer
By the Drook Daily desk · 4 min read
This is the third time in six weeks that AI-hardware stocks have swung on the same underlying question: does the compute buildout pencil out, or doesn't it? In June, a memory-glut scare wiped out over a trillion dollars in chip-sector value. Two weeks ago, SK Hynix's record $26.5B Nasdaq listing and a wave of efficiency-branded model launches put a bid back under the same names. This week, reports that SK Hynix is slowing its own HBM expansion sent them right back down — Super Micro −14.1%, Micron −11.1%, AMD −9.6% — while fifteen of the twenty-one AI-relevant stocks we track closed lower. Same story, third rerun, and the market still hasn't settled on an ending.
What's notable isn't the volatility itself — capex-sensitive cyclicals swing on sentiment all the time. It's what didn't move with it. Software and cloud names carrying no hardware capex on their books — Microsoft, Palantir, Salesforce — kept climbing through the same week chips fell hardest. And underneath all of it, capital allocators who write nine- and ten-figure checks kept writing them anyway: Abu Dhabi's MGX closed a $49B fund, Together AI raised $800M, Hyundai spent $325M to own its robotics supplier outright. None of that money is waiting for the public market's chip debate to resolve.
Public equity investors are re-litigating the same capex question monthly. The people who actually deploy capital into AI infrastructure — sovereign funds, strategic acquirers, GPU-cloud investors — have already moved on to writing the checks.
That split is the more useful signal than either week's direction. If the tape's swings told you something true about AI's economics, you'd expect private capital to hesitate alongside it. It hasn't. The more plausible read is that public-market sentiment overreacts in both directions to a genuinely uncertain multi-year capex cycle, while investors with longer horizons and direct diligence access underwrite through the noise. Two credible labs added to that uncertainty this week from a different angle: Moonshot's 2.8-trillion-parameter Kimi K3 bet that scale-at-any-cost still wins, while Thinking Machines' Inkling bet that a smaller, customizable model wins instead — and said so, unprompted, in its own launch post.
What this means for you. Don't let a single week's stock chart set your infrastructure or vendor strategy — it's reacting to the same unresolved question it reacted to in June and reacted to two weeks ago. Watch what's not swinging (this week: enterprise software) as much as what is. And treat the open-weight-vs-proprietary decision the way Moonshot and Thinking Machines are forcing it: as two live, unresolved bets, not a settled question with one right answer — pilot against both before you commit a platform to either.
06 — Off the Clock
Cool Stuff Going On
The fun, forward-looking launches — where AI quietly stopped being a chatbot and started doing things.
KIMIOpen Model
Kimi K3
Moonshot AI's 2.8-trillion-parameter open-weight model — the largest ever released — tops its own benchmark suite and leads blind coding evaluations against closed frontier models.
TMOpen Model
Inkling
Mira Murati's Thinking Machines ships its first model — built to be fine-tuned on the company's Tinker platform rather than to top a leaderboard.
WAICHumanoid
WAIC 2026's humanoid debuts
Shanghai's largest AI conference yet saw AgiBot unveil its A3 Ultra full-size humanoid and Unitree show off the transformable GD01 — four new product lines in one day.
URKOff the Clock
Robots throw hands in Shenzhen
Thirty-two teams of standardized EngineAI T800 humanoids competed in the "Ultimate Robot Knock-out Legend" for a $1.4M prize pool — the first robot combat league at this scale.
NEKOHealth Tech
Neko Health's $700M raise
The AI full-body scanning platform is expanding into the U.S. on fresh funding, betting that a 15-minute scan can catch what an annual physical misses.
UCHWearables
A cardiac AI patch that thinks on-device
University of Chicago researchers built a wearable that runs inference directly on the patch, delivering millisecond results for arrhythmia detection with no round trip to the cloud.
07 — Embodied AI
This Week in AI Robotics
Physical AI keeps consolidating — a full corporate buyout, the biggest robotics showcase of the year, and factories quietly hitting real production rates.
M&ABD
Hyundai buys out SoftBank, takes full ownership of Boston Dynamics
SoftBank exercised its 2021 put option; Hyundai is paying $325M for the remaining ~10% stake, valuing Boston Dynamics at $3.3B — unchanged since Hyundai's 2021 takeover. Hyundai targets its first Atlas factory deployment in Georgia by 2028, scaling toward 30,000 units a year. A car maker now owns its embodied-AI supplier outright — no partner to answer to on pace or direction.
ShowcaseWAIC
WAIC 2026 in Shanghai is the largest robotics showcase yet
AgiBot debuted four product lines Jul 18 — the A3 Ultra full-size humanoid, the X2 Edu platform, the G2 Max industrial robot, and the OmniHand 3 Ultra-M — while Unitree showed its transformable GD01. China's humanoid makers are now shipping product-line breadth, not single hero demos.
DeploymentBotQ
Figure's BotQ factory crosses 350 units delivered
The line is now producing 55+ humanoids a week with an expanding BMW Spartanburg pilot, moving Figure 03 onto real sequencing tasks rather than single-cell demos. Weekly production rate and cumulative units delivered are becoming the sector's real scoreboard.
CompetitionURK
Robot combat gets its first standardized league
Thirty-two teams fielded identical EngineAI T800 humanoids in Shenzhen's "Ultimate Robot Knock-out Legend," fighting for a 10M-yuan (~$1.4M) prize pool. Standardized hardware across every team turns this into a software and controls benchmark, not a spec-sheet contest.
Field dataBD
Boston Dynamics' electric Atlas logs its first real deployment data
Early field results are emerging from Hyundai and Google DeepMind industrial pilots, ahead of the 2028 Georgia factory target now that Hyundai owns the company outright. Full ownership removes the ambiguity of whose roadmap Atlas follows next.
◆ The number
$325M — what Hyundai paid to buy out SoftBank's remaining Boston Dynamics stake, at a valuation unchanged since 2021. For operators: a strategic acquirer just priced a name-brand robotics supplier flat over five years despite the humanoid funding boom around it — don't assume every robotics valuation is compounding at AI-hype speed.
08 — Your Sector
Industry Watch
AI moves in the verticals our readers run — read the one with your name on it.
▦
Construction
PartnershipMcCarthy Building Companies deployed Palantir's Artificial Intelligence Platform under a multimillion-dollar deal, building a connected AI operating system across the full project lifecycle.
The numbersAutomated estimating tools are hitting 85–90% accuracy against manually prepared bids, cutting a half-day process to minutes; pilot firms report 30–50% fewer admin hours from AI-driven field reports and dispatching.
The barrierA skills gap remains the top blocker to adoption — most firms still lack staff who can evaluate, deploy, and manage AI tools on the jobsite.
✚
Healthcare
OversightCMS stood up a new Office of Health Technology Products to oversee AI, interoperability, and digital-health tools across Medicare and Medicaid.
SavingsUnitedHealth (~$1B) and HCA (~$400M) still project AI-driven savings in 2026, largely from automating revenue-cycle work.
CapitalNeko Health raised $700M to expand its AI full-body scanning platform into the U.S. — one of the largest digital-health raises of the summer.
$
Finance
GuidanceThe U.S. Treasury released a shared AI Lexicon and a Financial Services AI Risk Management Framework for the sector.
Adoption77% of buy-side firms now run generative AI org-wide; broker research and earnings transcripts are the most-wanted machine-readable feeds.
In productionRevolut's proprietary model PRAGMA, built on unified customer data, lifted fraud detection 64.7% and credit-risk prediction 16%.
⇄
Logistics
Infrastructureproject44 launched LSP44, an AI-agent and API infrastructure layer already run by 9 of the world's 10 largest logistics providers.
AutonomyEnvoy AI launched Ellie Workforce, an operating system built for autonomous freight execution.
Reality checkRedwood Logistics argues exception-automation is "only the opening act" for AI in supply chains — the harder, higher-value work is still ahead.
09 — Looking Forward
The Week Ahead
Dates worth putting in front of your leadership team before they arrive.
July 27, 2026
Kimi K3's full open weights publish
Moonshot AI's 2.8-trillion-parameter model is live via API today; the full weight release under an MIT license is the moment self-hosting teams can actually put it through paces.
Late July 2026
Q2 earnings from the AI mega-caps
NVIDIA, Microsoft, Meta, Alphabet, and Amazon report into a live capex-confidence debate. Watch guidance closely: does this month's hardware selloff show up as caution, or do the hyperscalers reaffirm spend?
August 4, 2026
Public comment closes on the GSA's AI acquisition rule
The draft large-language-model-specific federal procurement regulation — already revised once after stakeholder pushback — closes for comment. If you sell into federal agencies, this is the window to weigh in.
2028
Hyundai's target for Atlas's first factory deployment
Now that Hyundai owns Boston Dynamics outright, its Georgia facility is the marker to watch for humanoid robotics moving from pilot to real production-line headcount.
“
It is not the most performant model available today, closed or open.
— Thinking Machines, in the launch post for Inkling, its first AI model
Forward this to one person who should be reading it.
The Drook Daily lands every Tuesday — ten minutes on exactly what changed in AI, written for the people who have to act on it.
