A weekly intelligence brief
Tech and Internet Desk · Weekly Dispatch
Claude Sonnet 5 ships as Anthropic and OpenAI both sit in the SEC quiet room. Chip stocks wobble on margin fears even as TSMC heads into a record quarter. Iran-linked hackers triple their hit rate on Israel. Jupiter moves into Leo and Mercury turns retrograde in the same week Wall Street starts pricing the first AI-lab IPOs.
Weekly Brief | Analyst Desk | 3 July 2026
Anthropic and OpenAI are both now sitting in the same regulatory quiet room, and the pace of product releases has not slowed while they wait. Anthropic shipped Claude Sonnet 5 on 30 June, its most agentic mid-tier model yet, performing close to flagship Opus 4.8 on many tasks and priced below the outgoing Sonnet 4.6 through the end of August. OpenAI filed its own confidential S-1 on 9 June, roughly a week behind Anthropic, at a reported valuation near 920 billion dollars, though the company has signalled the actual listing 'may be a while.' Both labs are now racing product cycles while lawyers assemble the paperwork that will, for the first time, force a frontier AI company to show its accounts to the public.
Chip stocks had a rough end to June even as the fundamentals stayed strong. Nvidia and Alphabet sat out a tech bounce on 26 June as investors worried aloud that soaring memory-chip prices could squeeze hyperscaler margins, and SoftBank, whose fortunes ride on Arm and its OpenAI stake, fell more than 5 percent in the same sell-off. Underneath the wobble, the physical build-out keeps accelerating: TSMC heads into its 16 July earnings call guided for 39 to 40.2 billion dollars in quarterly revenue, and full-year capital spending is being pushed to the top of its 52 to 56 billion dollar range. The gap between what investors fear this week and what the factories are actually doing this quarter is the story to watch.
Israel's cyber chief says hostile activity aimed at the country roughly tripled this year: about 4,800 incidents in June 2026 against roughly 1,600 in June 2025, with Iran-linked groups now sharing tools and methods rather than working alone. Critical infrastructure has held; smaller law and accounting firms have taken the worst hits, in some cases wiped entirely. Russia's own internet, meanwhile, keeps closing in on itself: more than 20 major domestic sites, including the state services portal and the country's biggest bank, now block visitors who arrive over a VPN, even as roughly 40 percent of Russians still use one daily to reach Telegram and WhatsApp.
Two long-run technology bets moved forward this week outside the AI headlines. Humanoid robots crossed from demo to shift work: Figure's factory in California is now building one robot per hour, a 24-fold jump in under four months, while its earlier model has helped assemble more than 30,000 BMW cars on a real production line. And fusion power took a regulatory step that matters more than any lab result: Tennessee became the first US state with its own rulebook for licensing a fusion plant, and Washington State handed Helion Energy the operating licence for its Orion facility. Nobody is running a commercial fusion grid yet, but for the first time governments are writing the paperwork as if someone will.
| Lane / Country | The one-line read |
|---|---|
| AI (global) | Claude Sonnet 5 ships 30 June. Both Anthropic and OpenAI now hold confidential S-1s at the SEC. Combined 2026 hyperscaler capex near 725 billion dollars, up 77% on 2025. |
| Semiconductors | TSMC guides Q2 revenue 39-40.2 billion dollars ahead of 16 July earnings. Chip stocks wobbled late June on margin fears. Huawei targets 600k Ascend 910C units; China HBM output still the ceiling. |
| Cyber warfare | Hostile incidents against Israel roughly tripled year on year, per the national cyber chief. North Korea crypto theft holds near 577 million dollars through April, 76% of all 2026 losses. |
| Splinternet | Russia: 20-plus major domestic sites now block VPN users outright. EU AI Act high-risk deadline confirmed deferred to Dec 2027 (final Council sign-off 29 June). New EU ban on AI-made intimate deepfakes. |
| United States | Anthropic and OpenAI both hold confidential S-1s. AI OVERWATCH Act still stuck in committee. Section 122 tariff surcharge set to lapse 24 July, a live variable for Israeli tech-goods trade. |
| China | Ascend 950PR in mass production since April; 950DT due Q4. CXMT HBM3 output ramping but yields near 50%. Cybersecurity Law amendments tighten data localisation from 1 January. |
| Israel | Cyber incidents up roughly 3x year on year. Startups raised 3.3 billion dollars in June alone; Cyera raised 600 million at a 12 billion valuation. US tariff deadline of 24 July still unresolved. |
| EU + Czechia | AI Act delay finalised 29 June by the Council. New ban on AI-generated non-consensual intimate imagery. Czech AI funding still thin in absolute terms but now near a third of all national VC. |
| Thailand | BOI-approved digital investment tops 29 billion dollars this year, led by TikTok data centers. TH-AI Passport contract dispute continues without formal findings. |
| Russia | Telegram and WhatsApp remain blocked without a VPN. Over 20 major domestic platforms now refuse VPN traffic. State messenger Max still shows signs of probing devices for VPN use. |
| Uzbekistan | IT Park continued its diaspora and export push with a Chicago meetup and a Poland business mission in June; Q1 service exports were 191.8 million dollars, up 38.5 million year on year. |
| Futurology | Fusion gets its first state licensing frameworks. Humanoid robots move from pilot to paid factory shifts. Starship static-fires toward its next flight. Neuralink moves toward high-volume implants. |
Snapshot as of 3 July 2026. Each lane and country is covered in full below.
FRONTIER MODELS / COMPUTE ARMS RACE / POWER
Anthropic released Claude Sonnet 5 on 30 June 2026. It is the new default model for free and paid users of claude.ai and is live in Claude Code, the Claude API, Cursor, VS Code and GitHub Copilot. Anthropic describes it as the most agentic Sonnet model built so far, meaning it is tuned to carry out long, multi-step tasks with less supervision, and it performs close to the flagship Opus 4.8 on many benchmarks despite being the cheaper, faster tier. It ships with a 1 million token context window (roughly 750,000 words of working memory in a single conversation) and is priced at introductory rates of 2 dollars per million input tokens and 10 dollars per million output tokens through 31 August, rising afterward to 3 and 15 dollars.
The release lands in a crowded field. OpenAI's GPT-5.6 introduced more predictable prompt caching for developers building always-on agents. Google's Gemini 3.5 Flash continues to compete on speed and cost rather than raw capability. Microsoft's own coding-focused models, unveiled in early June, are aimed squarely at reducing the company's reliance on OpenAI's models inside its own products. For a non-specialist: no single model wins every category any more. The competitive question has shifted from 'which model is smartest' to 'which model finishes a real multi-step job reliably and cheaply,' and that is a much harder thing to benchmark or to market.
Anthropic filed a confidential draft S-1 with the SEC on 1 June, four days after closing a 65 billion dollar Series H round at a 965 billion dollar valuation. OpenAI followed with its own confidential S-1 on 9 June, at a valuation reported near 920 billion dollars. Both filings are confidential, meaning the companies' financials are visible to regulators but not yet to the public; that changes only when each company sets a listing timeline and the filing goes public. OpenAI has signalled its actual listing could slip into 2027 as it works to secure investor demand at the trillion-dollar mark it wants.
To put the numbers in context: 965 billion and 920 billion dollars sit in the range of Visa's or Berkshire Hathaway's entire market value, for two companies that do not yet report public earnings. Whichever files first will set the template other AI labs get valued against for years. That is the real prize in the current race: not who lists first, but whose numbers become the market's reference point for what a frontier AI company is worth per dollar of revenue.
US technology media covers this as a two-company race to Wall Street. What that framing understates is how much both companies still depend on continuous fundraising rather than profit to fund the compute their models require. Neither Anthropic nor OpenAI has published audited annual profit figures. The S-1 process, once public, will be the first time either company's unit economics face outside scrutiny in a legally binding document, which is a different kind of pressure than an investor pitch deck.
The model-release news cycle, Sonnet 5 this week, GPT-5.6 the week before, keeps public attention on capability. Moving underneath: chip and cloud analysts spent the same week publicly worried about whether hyperscaler margins can absorb the memory-chip price increases driving up the cost of building the infrastructure these same models run on. The stock wobble on 26 June, when Nvidia and Alphabet sat out a broader market bounce, is a small but real signal that some investors are starting to price in a gap between AI spending and AI earnings.
Combined 2026 capital spending across Amazon, Alphabet, Microsoft and Meta is guided near 725 billion dollars, up 77 percent from about 410 billion in 2025. Amazon is spending roughly 200 billion dollars, up from 125 billion last year. Google is guiding 175 to 185 billion, up from 91 billion. Microsoft is planning 110 to 120 billion, up from 90 billion. Meta has guided to 115 to 135 billion, up from 72 billion. This is the largest single-year capital spending increase in the history of the technology industry, by a wide margin.
To say whether 725 billion dollars is a lot: the entire global semiconductor industry's annual revenue is roughly 600 billion dollars. The four hyperscalers, between them, are spending more than that just on the buildings, cooling and power infrastructure that house the chips, before counting the chips themselves. This is a bet on future AI revenue that independent analysts estimate is running well ahead of what AI products currently earn across all four companies combined.
Every major hyperscaler now has at least one nuclear power deal on the books for AI data-center capacity, and the combined commitment across announced projects sits above 9.8 gigawatts. One gigawatt is roughly the output of a large conventional power plant; what the sector has committed to is the equivalent of adding roughly ten large power plants to the grid over the next several years for AI alone. This week's late-June stock jitters over chip and memory prices sit on top of that structural fact: even fully funded, this build-out is gated by how fast new power can reach the grid, not by how fast anyone can write a cheque.
Capex headlines dominate. Less examined: the same companies also hold large government cloud contracts, which raises a fair question about which parts of the business are actually funding which. Anthropic and OpenAI's eventual public filings will be the first real window into whether frontier AI economics work at this level of spending, or whether the industry is relying on continued private fundraising to paper over the gap until public investors take over that role.
TSMC / EXPORT CONTROLS / CHINA SELF-SUFFICIENCY
TSMC reports second-quarter 2026 results on 16 July, guided for revenue of 39.0 to 40.2 billion dollars, roughly 10 percent higher than the prior quarter and about 32 percent higher than a year earlier, with a gross margin of 65.5 to 67.5 percent. May's monthly revenue alone rose 30.1 percent year on year to about 320.5 billion New Taiwan dollars. Full-year 2026 revenue is guided to grow above 30 percent in dollar terms, and full-year capital spending is being pushed toward the top of TSMC's 52 to 56 billion dollar range, the money that pays for new fabs and equipment. The company has also begun raising prices 5 to 10 percent on its most advanced processes, which make up about 75 percent of total revenue, a sign that demand still exceeds what TSMC can build.
To put these numbers in context: a modern TSMC chip packs tens of millions of transistors into an area the size of a fingernail, at a precision no other company matches at commercial volume. Nvidia's AI accelerators, which power most large models worldwide, come off TSMC lines. If TSMC's output slows, the global AI build-out slows with it; if TSMC can raise prices 5 to 10 percent without losing customers, that is the clearest available proof that AI chip demand has not cooled, whatever the stock market did in late June.
TSMC's dominance concentrates geopolitical risk in one place: Taiwan. The Q2 guidance number is a reliable read on whether the broader AI build-out is holding up; a miss here would be a far more credible signal of a slowdown than any single stock's daily move. The price increases are also a quiet answer to the margin-squeeze worry that rattled chip stocks on 26 June: TSMC is passing higher costs through to customers rather than absorbing them, which protects its own margins even if it tightens everyone else's.
Strong guidance and price increases absorbed most coverage. Moving quietly underneath: US export control policy toward China's chip-equipment supply chain continues to shift. The Commerce Department has renewed TSMC's annual approval to export chipmaking products to its mature-node China fabs (16 nanometre and larger, well behind the cutting edge), alongside similar approvals for Samsung and SK Hynix. That is a narrow, carefully bounded concession, not a reopening of advanced chip trade, and it is easy to conflate the two in casual reporting.
The core policy picture has not reversed since the Trump administration's January 2026 shift from an outright H200 export ban to case-by-case licensing with a 25 percent tariff: Nvidia CEO Jensen Huang said in March that the company had regulatory clearance from both Washington and Beijing to resume H200 sales into China, and orders have followed, though volumes clearing Chinese customs remain the harder number to verify. The AI OVERWATCH Act, which would give Congress a 30-day veto over chip export licences to adversary nations and a two-year ban on Blackwell-tier exports to China, passed its House committee 42-2 in January and has still not reached a full floor vote.
On the China side, Huawei's Ascend 950PR entered mass production in April, delivering roughly 2.9 times the inference performance of Nvidia's older H20 chip at about a quarter of the cost, and Huawei is targeting 600,000 units of its Ascend 910C this year, with up to 1.6 million total Ascend dies across the product line. A follow-on chip, the Ascend 950DT, aimed at the more demanding training workloads, is due in the fourth quarter. Huawei will showcase its next supernode cluster hardware at the WAIC conference in Shanghai from 17 to 20 July.
High bandwidth memory, the fast memory stacked directly onto AI chips, remains China's hardest bottleneck. CXMT, China's leading memory maker, only began mass-producing HBM3 in early 2026 and initial shipments to Huawei and other domestic chip designers are running at roughly 50 percent yield, meaning about half of what CXMT tries to produce is unusable. Estimates of China's 2026 HBM output vary widely, from around 2 million to roughly 7 million memory stacks depending on the analyst, but even the higher estimate is only enough to support chips roughly comparable to Nvidia's older H100, not its current Blackwell generation. Samsung, SK Hynix and Micron, all outside China, remain the only producers of the newest HBM generations at real volume.
Headline production targets (600,000 Ascend units, 1.6 million dies) get repeated without the caveat that Huawei does not publish verified production figures and outside estimates vary. The HBM yield problem rarely makes the same headlines as the unit targets, but it is the more important number: a chip with plentiful logic capacity and scarce, low-yield memory runs well below its spec-sheet performance in practice, and that gap is where China's self-sufficiency claims are weakest.
ISRAEL-IRAN / NORTH KOREA CRYPTO / STATE-SPONSORED OPERATIONS
Yossi Karadi, head of Israel's National Cyber Directorate, told the German newspaper Die Welt in reporting published 29 June that hostile cyber incidents against Israel reached roughly 4,800 in June 2026, against about 1,600 in June 2025, close to a threefold increase. Karadi said Iran-linked groups are increasingly coordinating and sharing attack methods rather than operating in isolation, and summarised the state of play bluntly: there is no ceasefire in cyberspace, even where there is one on the ground. Israel says critical infrastructure, power, water, major financial systems, has held so far. The firms taking the worst damage are smaller ones outside that shielded tier: law firms and accounting practices, some of which have had their systems wiped entirely rather than just breached.
To put 4,800 incidents in a month in context: that is over 150 attempted intrusions or attacks recorded per day against a country of about 9.8 million people, a hostile-activity density among the highest reported anywhere in the world. The shift Karadi describes, from lone-actor Iranian operations to coordinated, tool-sharing campaigns, mirrors the pattern security researchers have already documented among Russian-aligned ransomware groups: cheaper, more repeatable attacks scale faster than one-off skilled intrusions.
The gap between the official line, infrastructure is holding, and the practical experience of a wiped small accounting firm is not a contradiction; it reflects a triage choice. Israel's most resourced defences protect the assets whose failure would be catastrophic at national scale; smaller firms outside that perimeter absorb a disproportionate share of the damage. This pattern is not unique to Israel, but the scale of the reported increase, a near tripling in twelve months, is unusually steep even by the standards of an active conflict zone.
The 4,800-incident figure comes from a single named Israeli official quoted in one German outlet, not from an independent technical report with incident-by-incident detail. That does not make it false, Karadi holds the country's top cyber-defence post and has no obvious incentive to inflate the number, but it should be read as an official estimate rather than an audited count until a fuller report or corroborating source appears.
North Korean-linked hackers stole approximately 577 million dollars in cryptocurrency through April 2026, capturing 76 percent of all crypto hack losses recorded in that period, via just two incidents: a 285 million dollar breach of the Drift Protocol on 1 April, following three weeks of staged social engineering with Drift employees, and a 292 million dollar exploit of KelpDAO on 18 April that compromised internal network nodes and forced a bridge protocol's verifier to rely on poisoned data. Arbitrum's security council froze roughly 75 million dollars of the KelpDAO funds still on-chain using emergency powers.
For a country with an estimated GDP of roughly 18 billion dollars, 577 million dollars in four months is a meaningful state revenue stream, not petty theft, and North Korea's cumulative crypto theft since 2017 now exceeds 6 billion dollars. TRM Labs, the blockchain intelligence firm tracking these numbers, reports that North Korean operators increasingly use AI tools in the reconnaissance and social-engineering phases of these attacks, which lowers the cost of finding and manipulating human targets without needing more attackers.
RUSSIA RUNET / CHINA DATA RULES / EU AI ACT
Russia's sovereign internet law, in force since 1 March under Government Decree No. 1667, has kept escalating through the spring and into summer. Telegram availability without a VPN fell to roughly 5 percent by early April; WhatsApp was blocked two days after Telegram; YouTube remains throttled. Roskomnadzor had blocked 469 VPN services as of the last count in February. In mid-April, at the Ministry of Digital Development's instruction, more than 20 major Russian websites and platforms, including the Gosuslugi state services portal, Ozon, Wildberries, Yandex, Sberbank and VK, began refusing access to users connecting over a VPN, and reports since then describe similar denials spreading to banking and ride-sharing apps. Despite the Telegram block, its Russian user count fell only about 1.5 percent in the first quarter, holding above 94 million, because so many users simply moved behind a VPN.
The state's preferred alternative, Max, a messenger built on VK's platform and pushed into official Russian life by a June 2025 presidential decree, has not won meaningful voluntary adoption. A technical audit of its Android version found it probing users' devices for VPN activity and sending traffic to third-party servers to test which domains a device could reach. Roughly 40 percent of Russian internet users now rely on a VPN daily, an increase, not a decrease, since the crackdown began. A survey found 85 percent of Russian companies describe the effects of the internet restrictions on their operations as critical or significant.
Blocking a platform and demanding VPN-free access to state and banking sites are two different moves with the same underlying purpose: not simply hiding content, but forcing a choice that creates a legible record of who is trying to circumvent the system. The deep-packet-inspection infrastructure behind all of this, known as SORM, has been installed at every Russian internet provider since 2019; this year's blocking wave is the enforcement phase of infrastructure built over roughly seven years, not a sudden new capability.
Amendments to China's Cybersecurity Law took effect 1 January 2026, raising the maximum penalty to 10 million renminbi (roughly 1.4 million dollars), widening the law's reach to foreign companies operating outside China but affecting Chinese users, and folding AI governance into the same framework. Operators of what China defines as critical information infrastructure must store personal information and 'important data' collected inside China within the country, and any transfer abroad now requires a government security assessment from the Cyberspace Administration of China. That is a materially stricter approach than the EU's GDPR, which allows cross-border data flows under defined safeguards without requiring the data to stay put.
Separately, the Great Firewall's AI-driven upgrades from the past year, including detection layers for encrypted DNS traffic and heuristic identification of disguised VPN traffic, continue to push up the failure rate for VPN protocols that worked reliably inside China as recently as 2023. A leak of internal documents from Geedge Networks, a firm linked to a co-architect of the original Great Firewall, previously reported exports of equivalent 'secure gateway' technology to Pakistan, Myanmar, Ethiopia and Kazakhstan; none of the four governments has formally confirmed or denied the documents.
China's model treats data as a resource that must physically stay inside the country's legal reach, which is a structurally different bet than the EU's model of regulating what happens to data wherever it travels. Both are responses to the same underlying anxiety, that a company or government elsewhere might use citizens' data in ways the home government cannot see or stop, but China's version is far more restrictive on cross-border business, and the January amendments make it more so, not less, even as Washington and Beijing negotiate over chips.
The European Parliament granted final approval on 16 June to the Digital Omnibus amendments agreed in May, and the Council of the EU gave its own final green light on 29 June, formally confirming that compliance for high-risk AI systems under Annex III (systems used in hiring, credit scoring, benefits decisions, education and similar high-stakes contexts) is deferred from 2 August 2026 to 2 December 2027, a 16-month delay. High-risk systems that are safety components of regulated products get an additional 12-month delay, to August 2028. The same package added a new outright ban on AI systems that generate non-consensual sexual or intimate imagery of real people, including tools that digitally remove clothing from existing photos, effective from December this year.
None of this touches the fine structure already in force. Violations of the AI Act's banned practices, enforceable since February 2025, still carry fines up to 35 million euros or 7 percent of global annual turnover, whichever is higher, the harshest digital-regulation ceiling in the world. The Digital Services Act is separately live and enforcing: the European Commission's 120 million euro fine on X, confirmed in this cycle, remains one of the largest DSA penalties issued to date, for deceptive verification design and failing to give researchers promised data access. On 1 June, the Commission also appointed a 60-member Scientific Panel and a new Advisory Forum to support AI Act enforcement, and on 10 June it published a voluntary Code of Practice on transparency for AI-generated content.
ROBOTICS / FUSION / SPACE / BRAIN INTERFACES
Figure AI scaled production of its Figure 03 humanoid robot to one unit per hour at its BotQ factory, a 24-fold increase in throughput achieved in under 120 days, and says that pace is now sustained rather than a one-off burst. Its earlier Figure 02 model has already helped build more than 30,000 BMW X3 vehicles across a ten-month deployment at the Spartanburg, South Carolina plant, which makes BMW the clearest paying industrial customer for humanoid labour anywhere right now. Tesla's Optimus programme is running behind on that measure: its Gen 3 robot, with 22-degree-of-freedom hands and Tesla's AI5 chip, is targeting low-volume production at Fremont this summer, but Tesla has no announced external paying customer yet, and independent trackers currently score Figure well ahead of Tesla on real-world deployment traction.
To put 'one robot per hour' in context: that is a conventional automotive assembly-line cadence, not a lab demo cadence. The gap between the two companies is no longer about which robot looks more capable in a video; it is about which one has a factory customer willing to run it on a real line for ten months and report the output. That is a much higher bar, and only one of the two companies has cleared it so far.
Fusion power crossed from lab milestones into regulatory infrastructure this June. Tennessee became the first US state to adopt its own regulatory framework for licensing fusion machines, intended to prepare the ground for Type One Energy's proposed commercial fusion plant near Oak Ridge. Washington State's Department of Health granted Helion Energy the operating licences needed to run its Orion facility, confirming Helion has the staff, safety programme and physical plant in place to operate, not yet that Orion is delivering commercial grid power. The US Department of Energy approved Xcimer Energy's preconceptual design for its Athena laser-fusion plant, and a newer entrant, Inertia, raised 450 million dollars in Series A funding on 18 June with a stated goal of breaking ground on a grid-scale plant by 2030. The Department of Energy also published its Fusion Science and Technology Roadmap on 9 June, the first attempt to lay fusion's science, workforce and commercialisation priorities into one national strategy.
None of this is the same as fusion electricity reaching a home outlet. What changed is that regulators are now writing rules for a technology that, until recently, had no licensing pathway to write rules for. A licence is a bet that the technology is close enough to real that the paperwork is worth doing; it does not guarantee the plant works as designed.
SpaceX debuted on the Nasdaq on 12 June under the ticker SPCX at a valuation of roughly 1.77 trillion dollars, marking the moment the company's day-to-day launch cadence, still dominated by routine Falcon 9 Starlink missions, started trading as a public financial instrument rather than a private one. Starship's next-generation Version 3 vehicle completed a static fire of one of Ship 40's Raptor 3 engines on 26 June, a roughly 15-second full-duration burn, a step toward testing all six of that ship's engines and all 33 of its Super Heavy booster's engines ahead of a full flight. NASA's Artemis programme still depends on a version of Starship as the crewed lunar lander for Artemis 4, currently scheduled for late 2028.
Neuralink is moving toward what Elon Musk has called high-volume production of its brain-computer implant in 2026, alongside a streamlined, largely automated surgical procedure that threads electrodes through the dura, the brain's tough outer membrane, without needing to remove a section of it as earlier procedures did. More than a dozen people with severe paralysis now carry an implant, with the longest-running participant logging thousands of hours of continuous use controlling a cursor, playing games and posting online by thought alone. The company holds an FDA Breakthrough Device Designation specifically for speech restoration, a pathway that can speed up review for devices addressing conditions with no existing good treatment. The technology remains a clinical trial, not an approved consumer or even standard medical device; 'high-volume production' describes manufacturing capacity, not regulatory clearance.
Thailand's Board of Investment has approved roughly 29 billion dollars in digital infrastructure projects this year, headlined by ByteDance's 25 billion dollar TikTok data-center expansion across Bangkok, Samut Prakan and Chachoengsao, the last of which is emerging as Thailand's main data-center cluster alongside tenants including Gulf Development. Q1 2026 BOI investment applications topped 1.01 trillion baht (about 30.7 billion dollars) across 624 projects, with the digital sector alone accounting for 873.7 billion baht of that figure, data centers and cloud services doing most of the work. Combined with earlier commitments from Amazon Web Services, Microsoft (partnered with Charoen Pokphand Group and True Corporation) and Gulf Development's own 4.3 billion dollar five-year expansion, Thailand is now one of the most active data-center markets in Southeast Asia.
The structural risks are power and heat, not capital. Liquid cooling infrastructure at this scale is not yet widespread at Thai sites, and the power grid is already stretched during the dry season when hydroelectric output drops. Regulatory readiness and skilled-labour supply are secondary constraints the government has not squarely addressed. A parallel governance dispute continues: the Ministry of Digital Economy and Society has still not halted the 1.62 billion baht (about 45 million dollar) TH-AI Passport sovereign AI contract despite an opposition-filed anti-corruption investigation request; no findings have been issued either way.
Russia's tech sector keeps splitting on two tracks. The censorship track, the Telegram and WhatsApp blocks, the growing wall of domestic sites refusing VPN traffic, and the Max surveillance-messenger push described in Lane 4, represents the most aggressive enforcement of the country's internet isolation architecture since it was built. The domestic-AI track runs in parallel: Yandex's YandexGPT and Sberbank's GigaChat remain the main large language models in production, both constrained by chip access. Nvidia's advanced GPUs are under full export control to Russia, domestic chip production through Elbrus and MCST runs several generations behind, and the January 2026 US policy easing that opened a path for H200 sales to China does nothing for Russia, which remains under full-spectrum sanctions with no comparable carve-out.
The IT sector lost an estimated 50,000 to 100,000 tech workers after the 2022 invasion of Ukraine, though some have returned. The VPN crackdown is now beginning to affect the same tools that let Russian developers keep reaching GitHub, foreign cloud services and international professional networks after 2022, a slow-burn effect on domestic tech capacity that the blocking headlines do not fully capture.
Uzbekistan's IT Park reported first-quarter 2026 service exports of 191.8 million dollars, up 38.5 million dollars year on year, and kept pressing the export story abroad through June. IT Park Ventures backed three startups, NavAI, EduFuse and AzizAI, through a pitch competition tied to the GSMA M360 Eurasia conference, and separately put 1 million dollars into a 4 million dollar funding round for Ayan Capital, a London-based fintech founded by Uzbek entrepreneurs, its fourth investment through the fund. On 18 and 19 June, IT Park led a business mission to Poland, taking part in the ABSL Summit 2026 to build ties with European business-services and technology firms, following a similar diaspora-facing meetup held in Chicago.
The government's long-term target remains 5 billion dollars in annual IT exports by 2030, against a Q1 2026 annualised run rate of roughly 767 million dollars, a gap that requires close to 6.5 times growth in four years. The structural risks are unchanged: dependence on Russian IT workers who relocated after 2022 and whose status could shift under political pressure, and a Western-facing business model that could be squeezed if Russia-West tensions deepen.
Israeli startups raised 3.3 billion dollars in June 2026 alone, one of the strongest single months on record, led by cybersecurity firm Cyera's 600 million dollar round and networking company DriveNets' 410 million dollar raise. Fourteen Israeli cybersecurity companies were named among the 30 firms on the Rising in Cyber 2026 list, based on a survey of roughly 150 chief information security officers at large global enterprises, nearly half the list. Investors are concentrating on AI-native security companies built to protect the new generation of AI agents and workloads, the same trend line that shows up in the cyber-incident surge described in Lane 3: more attacks are generating more defence spending in a tight loop.
The unresolved variable is trade policy. The 15 percent US tariff on Israeli goods, in force since earlier this year, mostly bypasses tech because roughly 70 percent of Israel's tech exports to the US are services and software, which are not subject to goods tariffs. But Section 122, the trade authority underpinning part of the tariff structure, is set to expire around 24 July, which could drop remaining goods tariffs toward zero if it lapses without renewal, or could be extended if Washington and Jerusalem do not finish the broader trade agreement reportedly in advanced negotiation. Either outcome affects the roughly 30 percent of tech exports that are physical goods, chiefly machinery and industrial equipment, more than it affects software and services.
The AI Act delay confirmed this week (see Lane 4) gives every EU member state's enterprise sector the same 16-month runway, but Czechia's position is worth separating out. Czech AI startups raised roughly 37.6 million dollars in the first quarter of 2026, still the lowest AI funding density in Central and Eastern Europe in absolute terms, yet the AI sector now captures close to a third of all national venture capital, a sign the country's investors are concentrating rather than diversifying. Prague accounts for roughly 83 percent of the national startup sector's value, put at about 19.2 billion euros. A 2026 policy change raising the research and development tax allowance to 150 percent of eligible costs is aimed squarely at keeping that concentration growing rather than losing firms to larger Western European hubs.
The identity taking shape is a business-to-business 'serious software' hub: applied AI and enterprise tools, not consumer apps, are where Czech firms are finding traction, including fraud detection and document-verification companies serving banks and other regulated industries across the bloc. The same EU regulatory environment that raises compliance costs also levels the field against larger US and Chinese competitors bidding for the same regulated contracts, which is the quiet advantage Czech firms are counting on.
Hyperscaler capex begins to show productivity returns by 2028, easing the margin worries that rattled chip stocks in late June. Power supply expands fast enough through nuclear and, further out, early fusion licensing to keep pace with AI demand. TSMC's price increases hold and spread modestly through the supply chain without breaking demand. Anthropic and OpenAI both price their IPOs well, establishing a durable framework for valuing frontier AI labs that keeps investment flowing rather than freezing it. Humanoid robots move from single flagship deployments like Figure and BMW to a wider set of paying industrial customers. In this path, the countries that built data-center capacity early, Thailand, Malaysia, Poland, capture a disproportionate share of the economic upside.
AI capex keeps running far enough ahead of revenue that one or more hyperscalers cut guidance before the end of 2026, and the late-June chip-stock jitters turn out to have been an early warning rather than noise. One or both of the Anthropic and OpenAI IPOs disappoints or slips further, undermining the trillion-dollar valuation benchmark both companies are counting on. Power and memory supply constraints slow the build-out faster than nuclear or fusion capacity can arrive. Congress finally moves the AI OVERWATCH Act, creating chip-export uncertainty that freezes new orders. China closes the HBM yield gap faster than expected, undercutting the export-control strategy. A major state-linked cyberattack, plausibly stemming from the Israel-Iran or North Korea fronts already active this year, forces a regulatory reset that slows AI and cloud deployment broadly.
For readers who follow this desk's cycle lens, kept strictly to pattern recognition, not prediction. Three significant astrological shifts land in immediate succession around this week's close. Mercury turned retrograde on 29 June at 5 degrees Cancer, a period that runs through 23 July with a shadow phase into early August. Cancer governs domestic accumulation, data and the private sphere. The classic reading of Mercury retrograde is contract delays, systems audits, discovered errors in filings, and communications breakdowns. For this sector specifically, the window overlaps almost exactly with the confidential S-1 review period both Anthropic and OpenAI are sitting inside, exactly the kind of paperwork-heavy administrative process this transit historically makes slower and more error-prone.
Jupiter entered Leo on 30 June, its first time in that sign since 2014, beginning a roughly twelve-month transit. Leo governs spectacle, public performance, leadership and being seen. The pattern maps cleanly onto the IPO arc unfolding in Lane 1: two companies each worth close to a trillion dollars preparing their first public appearance on Wall Street is Jupiter-in-Leo energy in concentrated form. The shift from Jupiter's prior transit through Cancer, private accumulation, the Series H and late-stage rounds that built these companies' war chests, into Leo, the public stage and investor scrutiny that follows an S-1 going live, is a clean archetypal handoff from funding phase to performance phase.
The sun meets that same freshly arrived Jupiter in Leo at the end of July, building toward a Full Moon in Aquarius on 29 July that opposes both, conjunct Pluto: a culmination-and-transformation signature sitting almost exactly on top of the TSMC earnings call, the WAIC chip showcase and the Israel tariff deadline, all clustered in the same third week of July. Saturn and Neptune remain in their long conjunction in Aries, limits meeting dissolution, which continues to map onto this sector's core tension all year: enormous financial projections running into hard physical walls, power supply, memory yields, chip nodes, that no amount of capital can simply buy past.
Primary sources, regulator releases, company filings and native-language outlets listed by section. Where a publication has a known editorial position, it is noted.
Prepared by the News Feed analyst desk. Checked against company filings, regulator releases and native-language sources as of 3 July 2026. Where figures are estimates or unverified, this is noted above. Verify before acting.