Weekly Brief | Analyst Desk | 17 July 2026
The AI industry's argument with itself came to a head this week: is the spending justified by demand, or has it outrun it? TSMC, which makes almost all of the world's most advanced chips, reported second-quarter revenue of 40.2 billion dollars, up 33.7 percent on the year, and said its advanced-packaging lines are so full that they are now the thing limiting how fast customers can grow, not the other way round. That is a demand-side problem, the good kind to have. Yet the same week saw roughly 1.3 trillion dollars wiped off global semiconductor stocks, because investors are no longer asking whether AI chips sell, they are asking whether the money being spent to build them will ever be matched by revenue.
The money at stake keeps growing. The five biggest US cloud and AI companies, Microsoft, Alphabet, Amazon, Meta and Oracle, are on track to spend 660 to 690 billion dollars combined in 2026 on data centres and chips, nearly double 2025's total. To size that: it is roughly the entire annual economic output of a mid-sized country, spent in a single year on computing infrastructure, before it is clear the applications built on top of it generate matching revenue. Microsoft alone plans about 190 billion dollars, up 61 percent on the year. That capex-sustainability worry, not any single bad product, is what triggered this week's selloff.
Money is also moving toward the AI labs themselves. Anthropic is reportedly lining up meetings with Goldman Sachs, Morgan Stanley and JPMorgan for a possible initial public offering later this year, a step that follows its 65 billion dollar funding round in May at a 965 billion dollar valuation. Meanwhile the model race kept moving underneath the money: Meta launched an agentic coding model aimed at developers, and Mira Murati's Thinking Machines Lab shipped its first public model just nine months after founding, a fraction of the three to five years incumbents usually take, while candidly admitting it is not the strongest model on the market.
This brief runs five lanes: artificial intelligence, computer chips, cyber warfare, the splinternet (the slow break-up of one global internet into national ones, with its own rules for AI), and futurology. Every figure is put against a benchmark, every company claim is labelled as a claim, and single-source reports are flagged as such rather than treated as confirmed fact.
Scoreboard: the five lanes
| Lane | Where it stands right now |
|---|
| Artificial intelligence | Meta and Thinking Machines Lab both shipped new models; Gemini 3.5 Pro slipped and hit Alphabet stock. Big Tech capex heading for 660 to 690 billion dollars in 2026. Anthropic lined up banks for a possible IPO. |
| Computer chips | TSMC posted record revenue and says packaging capacity, not demand, is the limit. Nvidia's China chip shipments remain a trickle even though billions in licences are approved. About 1.3 trillion dollars was wiped off global chip stocks. |
| Cyber warfare | The EU and UK jointly sanctioned Russia's FSB hacking unit for the first time, tied to an attack that threatened Poland's power grid. A single security firm reports the first fully autonomous AI-run ransomware attack. |
| The splinternet | Brussels proposed an EU-wide minimum social-media age and an anonymous age-verification app. The EU AI Act's next hard deadline is 2 August. Thailand's AI-access scheme stays mired in a conflict-of-interest dispute. |
| Futurology | A US fusion startup converted plasma directly into electricity on a lab bench, well short of grid power. Humanoid-robot IPOs kept coming. SpaceX scrubbed a Starship test flight after two engines failed to ignite. |
As of 17 July 2026. Vendor performance claims and single-source reports are flagged as such throughout, not treated as confirmed fact.
Artificial intelligence
New models, and one honest admission
Meta launched Muse Spark 1.1 around 9 to 10 July, an agentic coding model with an API aimed squarely at developers who currently use OpenAI or Anthropic's tools, confirmed across multiple outlets. Five days later, Thinking Machines Lab, the startup founded by OpenAI's former chief technology officer Mira Murati, released its first public model, Inkling. It is open-weight, built as a mixture-of-experts design (975 billion parameters in total, but only 41 billion, about 4 percent, actually run for any single answer, which is what keeps it affordable to operate) and trained on 45 trillion multimodal tokens, a training set on the scale used for today's frontier systems. The company itself says Inkling is not the single strongest model available today, open or closed, a rare piece of vendor honesty worth crediting rather than a marketing overclaim. The real story is speed: about nine months from founding to a shipped product, against three to five years for Anthropic or OpenAI to reach comparable milestones. A reported 50 billion dollar fundraise for the company had stalled as of January 2026.
Two claims to treat with care, one that moved a stock
Moonshot AI, a Chinese lab, released Kimi K3 on 16 July, a 2.8 trillion parameter open mixture-of-experts model, about three times the size of Inkling by parameter count. That figure comes from a single aggregator source, so treat the headline specs as reported rather than independently verified. PrismML's Bonsai 27B, released 14 July, is small enough to run on iPhone-class hardware at a claimed 90 percent of full-precision performance; both the model and the 90 percent figure are single-source and the percentage is the vendor's own claim, unreplicated. By contrast, Alphabet's problem was real enough to hit the share price: reports on 16 July that its flagship Gemini 3.5 Pro is behind schedule sent Alphabet stock down, a market reaction that confirms the delay is being taken seriously by investors, whatever the eventual release date turns out to be. Separately, benchmark scores circulating for Gemini 3.1 Pro (94.3 percent) and GPT-5.4-Pro (94.4 percent) on GPQA, a hard science question test used as one yardstick of model capability, come from aggregator sites and are not independently replicated; treat them as rumour-tier until a lab or independent evaluator confirms them.
The money behind the models
The five biggest US cloud and AI companies are on track to spend 660 to 690 billion dollars combined in 2026 on data centres and chips, nearly double 2025's total; Microsoft alone plans about 190 billion dollars, up 61 percent on the year. Anthropic is reportedly meeting Goldman Sachs, Morgan Stanley and JPMorgan about a possible initial public offering later this year. That follows Anthropic's 65 billion dollar Series H round in May, at a 965 billion dollar valuation that overtook OpenAI's 852 billion dollar valuation from March; Anthropic says its own run-rate revenue crossed 47 billion dollars in May 2026, up from a 1 billion dollar annualized rate at the start of 2025, a company-reported figure worth noting as such even though it has been widely repeated. Israeli tech investment also climbed, with 8.4 to 8.6 billion dollars raised in the first half of 2026, up 45 to 52 percent on the year, and cybersecurity investment more than doubling; recent rounds include Cylake's 45 million dollar seed round from Palo Alto Networks founder Nir Zuk, Ocean's 20 million dollar Series A for AI-driven email-attack defence, and Native's 31 million dollar Series A for multi-cloud security.
Computer chips
TSMC: record revenue, and demand outrunning even record spending
TSMC, the Taiwanese company that makes almost all of the world's most advanced chips, reported second-quarter revenue of 40.2 billion dollars on 16 July, up 33.7 percent on the year, with high-performance computing and AI now 66 percent of total sales, up from 55 percent a year earlier: two-thirds of the world's leading chipmaker's business is now AI chips, versus roughly half a year ago. The company raised full-year 2026 revenue guidance to slightly above 40 percent growth and lifted planned 2026 capital spending to 60 to 64 billion dollars, on top of a separate 100 billion dollar commitment to build multiple new plants in the US. Chief executive C.C. Wei said advanced-packaging capacity is so tight that it is now the thing limiting how fast customers can grow, meaning TSMC cannot physically build AI chips fast enough even at record spending. That is the clearest signal available that demand for AI chips, not supply of orders, remains the binding constraint.
A narrow, closely watched trickle to China
Nvidia's H200 chip began shipping to China in July after the US Commerce Department approved export licences for about 10 Chinese firms, including Alibaba, Tencent, ByteDance and JD.com, each capped at 75,000 chips. But Commerce official Jeffrey Kessler told Congress on 14 July that actual shipments so far are minimal and very few, even though 10 billion dollars in licences have been approved, meaning the headline approval figure is far bigger than what is actually moving across the border. Nvidia separately tightened its own compliance checks on buyers in Singapore, Malaysia and Japan on 14 July, reportedly cutting its approved Asia buyer list roughly in half and sending staff to visit customer data centres, to stop chips being rerouted into China through shell buyers. Read together, Washington is allowing a narrow, closely vetted trickle of mid-tier chips into China while working to plug every route that would let those chips, or the newer Blackwell chips that need a licence for any China or Macau transfer, leak further into China's AI buildout.
China's self-sufficiency push, and the wall it keeps hitting
Beijing is targeting 70 percent domestic silicon wafer use by the end of 2026 and 80 percent total chip self-sufficiency by 2030, up from about 33 percent in 2024, more than doubling its self-sufficiency share in six years under the plan, plus a new mandate requiring 50 percent domestically made equipment for any new fab capacity. The wall it keeps hitting is advanced lithography: SMIC, China's leading chipmaker, still depends on ASML's deep-ultraviolet tools to produce 7 nanometre chips, because the Dutch government's export ban on the more advanced extreme-ultraviolet machines remains in place. A prototype domestic EUV machine is in testing with a 2028 target for functional chips; outside analysts see 2030 as more realistic.
The 1.3 trillion dollar selloff
Samsung reported a record second-quarter 2026 operating profit of about 58 billion dollars on 7 July, 19 times its result a year earlier, and its stock still fell 7 to 10 percent, wiping out more than 100 billion dollars in market value, because investors read even that record result as confirmation that AI memory-chip demand cannot keep pace with the industry's capital spending. The reaction cascaded into roughly 1.3 trillion dollars wiped off global semiconductor market value over the following week, hitting Intel (down 10 to 21 percent across sessions), AMD (down 8 percent), Micron, Applied Materials and Lam Research. The plain read: even a company posting its best quarter ever can see its stock crash if investors doubt the industry's whole spending bet is sustainable. Intel reports earnings on 23 July and AMD on 4 August; neither reported this week.
Cyber warfare
A first: the EU and UK jointly sanction Russia's FSB
On 13 July the EU and UK jointly sanctioned Russia's cyber apparatus for the first time. The EU listed 9 individuals and 4 entities; the UK listed 24. Both formally attributed a cyber-espionage and sabotage campaign dating back to 2010 to FSB "Centre 16", the unit behind the group known as Turla, Secret Blizzard or Waterbug, with named targets including France, Germany, Poland, Cyprus, the Netherlands, Austria, Slovakia, Romania and Finland. The trigger was a December 2025 attack on Poland's energy grid that could have cut power to about 500,000 people in winter, roughly the population of a mid-sized city losing electricity in freezing weather. UK Foreign Secretary Yvette Cooper said the sanctions target the networks propping up Russian state aggression.
An overdue patch, and a signed driver turned against its own maker
CISA, the NSA and international partners are still urging emergency patching of Cisco Catalyst SD-WAN systems this week. Several vulnerabilities, including CVE-2026-20262 and CVE-2026-20245, sit in CISA's Known Exploited Vulnerabilities catalog with patch deadlines already about seven days overdue. It matters because SD-WAN edge devices carry privileged network access into government and critical-infrastructure networks. Separately, ransomware researchers reported on 9 to 10 July that a strain called "GodDamn" (a rebrand of the Beast and Monster ransomware lineage, according to Symantec) uses a Microsoft-signed kernel driver, nicknamed "PoisonX", to blind antivirus and endpoint-detection tools at the kernel level. The driver's Microsoft signature was obtained by misrepresenting its function during the signing process, an abuse of supply-chain trust rather than a technical exploit, confirmed by multiple independent security vendors.
A single-source claim worth watching, and two breaches
Researchers at security firm Sysdig say they have documented the first fully autonomous, AI-agent-run ransomware attack, which they call "JadePuffer": an AI agent reportedly carried out reconnaissance, credential theft, privilege escalation and file encryption with no human in the loop, entering through a vulnerability in the Langflow platform (CVE-2025-3248), and in one case self-corrected a failed login attempt in 31 seconds. This is genuinely significant if it holds up, but it is a single-source report from one security firm, not yet independently replicated by a second, so treat it as an early claim rather than settled fact. On breaches: AssuranceAmerica, a US auto insurer, disclosed that a phishing compromise detected in March exposed the driver's licence, policy and claims data of 6,998,886 people, nearly 7 million, more than the combined population of several US states. Deutsche Bank confirmed a breach at a third-party German service provider after a group calling itself "Unsafe" claimed to have leaked employee emails, password hashes and addresses; the bank denies its internal network was touched, so the scope of what was actually exposed is disputed between the bank and the attacker.
The splinternet
Brussels moves on children online, and the AI Act's next deadline
On 13 July, European Commission President Ursula von der Leyen proposed a three-part child-safety plan: mandatory accountability for platforms, an EU-run anonymous age-verification app, and a legally mandated minimum social-media age, with an expert panel recommending 13 with time-limited supervised access below that. It follows Australia, the UK, France, Greece and Spain, which have already enacted their own restrictions, and Malaysia, which banned social-media accounts for under-16s on platforms with more than 8 million users in June, the fourth country worldwide to do so, with fines of up to 1 million ringgit, roughly 210,000 to 230,000 dollars, for platforms and no penalty for parents; critics say the Malaysian rule risks pushing children onto unregulated platforms instead. Legislative proposals for the EU plan are expected after summer 2026. On the AI Act itself, chatbot-disclosure rules became legally binding on 10 July, so any AI chatbot serving EU users must now say it is AI, and the Digital Omnibus package, approved 9 July, pushed high-risk system obligations (biometrics, critical infrastructure, employment, border control) back to 2 December 2027 and product-embedded AI rules (lifts, toys) to 2 August 2028. The next hard deadline is 2 August 2026, when the Commission gains fining power over general-purpose AI providers and the Article 50 transparency and deepfake-labelling rules go live; regulators in Germany, France and the Netherlands have already opened first investigations, including into a CV-screening AI vendor in Germany. Czechia is separately drafting its own "Adaptation Act" to implement the EU law, naming its telecom regulator, central bank and data-protection office as enforcers, though the bill remains a draft.
Thailand's data-centre boom, and its unresolved AI-passport dispute
Thailand's Board of Investment approved roughly 29 billion dollars in new data-centre and digital-infrastructure investment, led by a TikTok Thailand data-centre project, part of a deliberate push to become ASEAN's data-centre hub. Set against that is a live political dispute: a 1.62 billion baht, roughly 45 million dollar, government scheme to give 5 million Thais a year of free premium AI access, funded through an extra-budgetary Digital Economy Fund rather than the normal budget process. Digital Minister Chaichanok Chidchob has admitted personally knowing a figure connected to the winning contractor and has refused opposition demands, from the People's Party and Democrat Party, to suspend the project pending an anti-corruption probe. This week negotiators are discussing shifting to a pay-per-active-user model to defuse the dispute, with a possible public launch by the end of July.
Rules in force in Russia since 1 March 2026 give the regulator Roskomnadzor authority to centrally control network traffic and disconnect the Russian internet segment from the global network over what it calls threats to stability. A whitelist regime covering government services, VKontakte, Yandex and more than 500 approved services is active across most regions, and by mid-2026 fewer than 10 percent of connection attempts outside that whitelist reportedly succeed. Telegram, used by more than 95 million Russians, has faced large-scale disruption, and Moscow and St Petersburg have seen partial mobile-internet shutdowns; regional and national shutdown incidents reportedly rose from 655 a month in June 2025 to 2,099 in July 2025 alone, already exceeding the prior full year's global total. Those figures come from OSW, a Polish state-linked think tank, and the Atlantic Council, both with a Western-institutional view of an adversarial government, so treat the magnitudes as best-available estimates rather than official Russian data. China, meanwhile, keeps exporting the tools of a controlled internet: at least 11 African countries have collectively spent more than 2 billion dollars on Chinese AI-powered surveillance systems, facial recognition and smart-city monitoring from Huawei, ZTE, Hikvision and Dahua, financed partly through Chinese state-bank loans conditioned on buying Chinese technology. Roughly 63 countries worldwide now have some Chinese-sourced AI surveillance infrastructure, and the reported pattern of misuse is that it is deployed more often to monitor political activists and opposition than to reduce crime, an investigative finding from Rest of World corroborated by the Atlantic Council and the South China Morning Post. Elsewhere, Uzbekistan is pitching itself as a coming regional tech hub through a 117 million dollar project to build IT parks across six cities, with the first Namangan facility expected to create about 2,000 jobs and 30 million dollars in exports, and a separate Fergana site aiming to train 3,000 specialists a year; IT Park member export revenue reportedly hit 191.8 million dollars in the first quarter of 2026 alone. Those figures come from Uzbek state and trade press covering a government-branded initiative, so treat them as officially reported rather than independently audited.
Futurology
Fusion: a lab-bench milestone, not grid power
Realta Fusion, a spinout from the University of Wisconsin-Madison, announced what it says is the first private-sector demonstration of converting fusion plasma directly into electricity, with no steam turbine involved. Tested on 19 June and still the dominant fusion story this week, the device used a direct energy converter on a magnetic-mirror machine to draw several amperes at about 100 volts, enough to light a few bulbs, a household-scale amount of power, nowhere near what a grid connection would require. Chief executive Kieran Furlong described it as an early practical application of the method on a working fusion device by a private company. Most fusion approaches plan to boil water and spin a turbine, so a working direct-conversion approach could be more efficient over the long run if it scales, but this remains a lab-bench proof of a physics principle, not grid power and not net-positive fusion energy; the "commercial" framing in the company's own announcement is promotional language to read with that caveat attached. Separately, General Fusion became the first publicly listed fusion company on 13 July, a financial and structural milestone for the sector, not a technical one.
Brains, robots and a scrubbed rocket
A "double neural bypass" implant restored movement and touch sensation in a quadriplegic patient, Keith Thomas, according to results reported on 16 July; it remains research-stage, not a commercial product. China's Neuracle reportedly performed the world's first surgical implant of a commercial brain-computer interface, an 8-electrode, coin-sized device called NEO, in a patient with a spinal-cord injury, a marker of China moving this hardware from the lab toward a commercial medical device ahead of some Western competitors; the claim is single-source, so treat "commercial first" as unverified against any global registry for now. In robotics, Agility Robotics is going public in a deal valuing the company at about 2.5 billion dollars and raising more than 620 million dollars, reportedly the largest capital raise in humanoid-robotics history, though its own chief executive is explicitly not promising a robot in consumer homes soon, a useful reality check on the sector. 1X Technologies upgraded the hand on its Neo humanoid to 25 degrees of freedom, aiming to match human dexterity, with the first household units due to ship this year; the claim that it can match or surpass human capability is 1X's own marketing, pending independent testing. In China, humanoid startups including LimX Dynamics, raising 200 million dollars ahead of a planned listing, are rushing toward IPOs, and one startup reports 13,000 pre-orders for a 17,600 dollar humanoid; China's stated national target is more than 10,000 humanoid deployments by the end of 2026. Pre-orders are not deliveries, and these figures are vendor-reported, not independently confirmed. Finally, SpaceX scrubbed its Starship Flight 13 test on 16 July at Starbase, Texas, after two Raptor engines on the Super Heavy booster failed to ignite before launch, triggering an automatic abort. It was the first Starship attempt since SpaceX's IPO and the debut of the new "V3" variant, carrying 20 next-generation Starlink satellites as a suborbital test payload; a relaunch is expected early next week. It matters for the long view because Starship is the vehicle NASA is counting on for the next crewed Moon landing.
The cycle view
Strict pattern recognition, not prediction. Saturn and Neptune's slow conjunction in early Aries, still in effect this week as it was last, reads in a tech frame as hard physical limits (a packaging line that cannot keep up, a lithography wall China cannot yet build past) running into hype and blurred claims (a lab-bench fusion demo framed as commercial, an AI ransomware attack confirmed by only one firm). Jupiter in Leo continues to favour bold, attention-seeking moves, which fits a week of new model launches, IPO plans and record-breaking capital raises. The steadier pattern underneath, as most weeks, is unglamorous: packaging capacity, export licences and regulatory deadlines are what actually decide who wins, far more than any single announcement.
Where this is heading
If capacity keeps the buildout justified
TSMC's packaging constraint turns out to be the real story: demand for AI chips keeps outrunning even record capital spending, this week's 1.3 trillion dollar chip selloff looks like a bout of nerves rather than the start of a correction, and the roughly 675 billion dollar Big Tech capex figure holds or grows. Anthropic's IPO process advances, the EU AI Act's 2 August deadline beds in without driving developers out of Europe, and humanoid-robot and fusion companies keep raising money on the promise that the hardware will eventually catch up to the pitch.
If the sustainability worry proves right
The selloff was the first crack, not a one-off: further earnings from Intel on 23 July or AMD on 4 August read as confirmation that AI spending has outrun AI revenue, the chip rout deepens, and IPO plans from Anthropic, Agility Robotics and others meet a colder market than they were built for. Nvidia's narrow, tightly policed trickle of chips into China stays throttled rather than widening, state-linked hacking campaigns like the one behind the Poland grid attack keep testing critical infrastructure regardless of the sanctions, and the internet keeps fracturing along national lines whichever way the capex bet goes, because that trend runs on its own political logic, not on chip demand.
Dates to watch
- 23 July Intel's second-quarter earnings, the first major chipmaker report since the 1.3 trillion dollar semiconductor selloff, a read on whether the worry was overdone.
- 2 August The EU AI Act's next hard deadline: the European Commission gains fining power over general-purpose AI providers, and the Article 50 deepfake-labelling rules take effect.
- 4 August AMD's second-quarter earnings, a second data point on chipmaker health after Intel reports.
- End of July A possible public launch of Thailand's AI-passport scheme under a revised pay-per-active-user model, if negotiators land on it in time.
- Ongoing SpaceX's Starship relaunch attempt, expected within days; Anthropic's IPO preparations; and China's progress toward its 70 percent domestic chip-wafer target by the end of 2026.
How sure we are
- Artificial intelligence The Muse Spark, Inkling and Gemini 3.5 Pro delay stories, and the Big Tech capex and Anthropic IPO figures, are confirmed across multiple outlets or official company statements. Kimi K3's specifications and Bonsai 27B's 90 percent performance claim are single-source and vendor-reported respectively, unverified. GPQA benchmark scores circulating for Gemini 3.1 Pro and GPT-5.4-Pro are aggregator-sourced rumour, not independently replicated.
- Computer chips TSMC's results are verified against its own earnings call transcript. Nvidia's China shipment numbers come from official Congressional testimony. Samsung's results and the broader semiconductor selloff are extensively cross-reported. China's self-sufficiency targets and EUV timeline are cross-referenced but remain forward-looking projections, not delivered outcomes.
- Cyber warfare The EU and UK sanctions on Russia's FSB are verified against the EU Council's own statement, and the Cisco SD-WAN advisories are official CISA alerts. The AssuranceAmerica breach is confirmed. The JadePuffer autonomous-ransomware claim comes from one security firm, Sysdig, and has not been independently replicated; the Deutsche Bank breach is confirmed but its scope is disputed between the bank and the attacker.
- The splinternet The EU's child-safety proposal and AI Act deadlines are confirmed via the European Commission's own materials. Thailand's data-centre investment and AI-passport dispute are extensively covered by Thai press. Russia's internet-isolation figures come from OSW and the Atlantic Council, both with a Western-institutional lens on an adversarial government, so treat magnitudes as best-available estimates. Uzbekistan's figures are officially reported by state and trade press, not independently audited, and the China surveillance-export figure is investigative journalism corroborated by two other outlets but not government-confirmed.
- Futurology Realta Fusion's direct-electricity demonstration is real but its "commercial" framing is company promotion; it is not grid power. The double neural bypass result is confirmed by STAT News; China's Neuracle "world's first commercial" BCI claim is single-source and unverified against any registry. Agility Robotics' IPO is confirmed. 1X's dexterity claims and China's humanoid pre-order figures are vendor marketing, not independently tested or delivered.
Sources
Official statements, wire services and specialist outlets; grouped by lane. Aggregator and content-farm sources were excluded except where explicitly flagged as such, and unverifiable claims are labelled in the text rather than presented as fact.
Artificial intelligence
Computer chips
Cyber warfare
The splinternet
Futurology
Plain-language glossary
The technology terms used in this brief, explained for a general reader.
- Mixture-of-experts. A way of building an AI model so that only a small slice of it actually runs for any single question, which keeps it cheaper to operate. Thinking Machines Lab's Inkling uses 41 billion of its 975 billion total parameters per answer, about 4 percent.
- Capital spending (capex). The money a company spends on long-lived physical assets, here mostly AI data centres and the chips inside them. The 660 to 690 billion dollars the five biggest US cloud companies plan for 2026 is the clearest measure of how large a bet Big Tech is placing on AI.
- Advanced packaging (CoWoS). The process of assembling multiple chip components into one finished AI processor. TSMC says this step, not chip manufacturing itself, is now the bottleneck limiting how many AI chips it can deliver.
- Extreme-ultraviolet (EUV) lithography. The most advanced machinery for printing circuits onto a chip, made only by the Dutch firm ASML and subject to a Western export ban on sales to China. It is the main technical wall standing between China and full chip self-sufficiency.
- GPQA. A test of hard graduate-level science questions used as one yardstick for how capable an AI model is. Scores that come only from company or aggregator reporting, without independent testing, should be treated as unverified.
- Ransomware. Malicious software that locks or steals a victim's files and demands payment to release them. One reported case this week allegedly ran with an AI agent handling the entire attack, with no human involved, a claim from a single security firm not yet independently confirmed.
- The splinternet. The slow break-up of one shared global internet into separate national internets, each with its own rules, censorship and, increasingly, its own AI regime. Russia's isolated network, with a whitelist of approved services, is the sharpest current example.
- EU AI Act. The European Union's law regulating artificial intelligence, the first broad one of its kind. Its chatbot-disclosure rules are already binding, and its next major deadline, 2 August 2026, gives regulators the power to fine general-purpose AI providers.
- Brain-computer interface. An implant or device that lets a person control a computer, or regain sensation, using signals from their brain, being tested by several companies for people with paralysis. Most results described here are research-stage, not commercial products.
Prepared by the News Feed analyst desk. Verified against official statements, wire services and specialist outlets as of 17 July 2026. Vendor and company performance claims are labelled as such. Not investment advice.
