Briefing Desk
Latest Briefings
Concise reporting on the developments shaping internet governance and infrastructure. Browse each area for recent news, context and watchpoints.
Woodburn council weighs data-centre moratorium
A moratorium would affect projects only after Woodburn defines which permit actions city staff must stop and from what date.
Mitsubishi's August target does not yet restore its quake-hit chip output
Mitsubishi Electric says two semiconductor factories halted after the Kumamoto earthquake should move back to normal operations during August. The timetable narrows the disruption, but it does not disclose current output, lost production or when customers will again receive normal volumes.
NTT's OptQC pact adds capital to a roadmap, not a million-qubit machine
NTT and OptQC have turned a research relationship into a capital and business alliance, with NTT planning to invest in the University of Tokyo spin-out. The agreement creates a longer operating horizon for photonic quantum computing, but the price, stake and funding date remain undisclosed, while the flagship machine is still a 2030 target.
Twilio sees Safaricom Kenya receipt recovery after 57 minutes, but the cause stays undisclosed
Twilio says SMS delivery receipts on a route to Safaricom subscribers in Kenya recovered after a 57-minute public incident window. The underlying messages may have succeeded while their acknowledgements arrived late. Service is being monitored, not declared resolved, and the operator has not explained the cause or quantified the affected traffic.
Zoom’s status object kept Meetings operational while reporting degraded quality
Zoom says a subset of users in its US region experienced degraded meeting quality for ten minutes on 2 August. The incident record appeared about 48 minutes after that interval ended, pairing an operational component state with a narrative account of impaired experience.
Michigan’s nine water-system reports widen the cyber incident without showing unsafe water
Michigan says nine water systems reported malicious activity consistent with a federal alert. The systems kept operating safely and local teams addressed the issues, leaving a serious operational-technology investigation but no disclosed public-health impact or named culprit.
Twilio saw recovery on a subset of SMS routes to Safaricom, but had not closed the incident
Twilio reported delays and failures from some alphanumeric sender IDs and long codes to Safaricom subscribers in Kenya. It identified an undisclosed cause, observed recovery after 94 minutes and was still monitoring stability at the fixed cutoff.
EU AI transparency duties take effect with separate obligations for providers and deployers
Article 50 of the EU AI Act now requires in-scope systems to disclose AI interaction, embed detectable machine-readable markings in synthetic output, and label specified biometric, deepfake and public-interest uses. The duties are binding; signing the supporting code remains voluntary.
Zoom disclosed a 76-minute Marketplace and AI Companion degradation only after it had ended
Zoom says users experienced degraded service across Marketplace and AI Companion from 00:00 to 01:16 UTC on 2 August. The incident record appeared roughly 29 minutes later, naming two service families but no affected function, cause, scale or geography.
Cloudflare’s 15-minute Istanbul network incident left a precise window but almost no diagnosis
Cloudflare says elevated latency and connection errors affected Istanbul for 15 minutes on 1 August. The unusually exact interval is the firmest part of a record that identifies no product, route, cause or scale—and whose only visible notice appeared after the reported impact had ended.
GitHub restored GPT-5.6 Luna after 74 minutes, but the incident exposes a model-routing dependency
GPT-5.6 Luna became degraded across GitHub Copilot products and IDE surfaces because of an unnamed upstream model provider on 1 August. GitHub offered a practical escape route—choose another model or Auto—but published no measure of how many requests failed or what functional trade-off the substitution imposed.
Twilio’s Globe Philippines incident delayed the receipt signal, not necessarily the SMS
For 2 hours and 42 minutes on 1 August, Twilio reported delayed SMS delivery receipts from its phone numbers to Globe Telecom subscribers in the Philippines. Messages could still succeed. That distinction matters because applications often treat the receipt as a control signal even when the underlying transport outcome is different.
Cloudflare’s RealtimeKit join failures lasted 65 minutes, while its incident metadata says otherwise
Cloudflare says some RealtimeKit users faced slow socket connections and failed meeting joins from 13:10 to 14:15 UTC on 1 August. Yet the incident object records both creation and resolution at 13:10:30. The narrative is operationally useful; the metadata cannot describe the same impact window.
GitHub took 17 minutes to name Fable 5 after upstream model errors began
GitHub’s second public Copilot model-provider incident on 1 August began with generic reports of upstream errors and only later identified Fable 5. Mitigation followed 28 seconds after that model-specific disclosure. The chronology shows fast recovery once the route was named, but not whether discovery itself or public attribution consumed the preceding 17 minutes.
Cloudflare’s IAD 5XX episode lasted 256 minutes but was disclosed in one retrospective update
Cloudflare says an increased level of HTTP 5XX errors affected Ashburn, US (IAD), from 18:45 to 23:01 UTC on 31 July. Yet the public incident contains only one visible update, timestamped more than two hours after that interval ended. The fault is closed; the operational sequence, affected product and scale remain largely invisible.
Twilio’s Claro Colombia short-code delay was still under investigation at the 20-minute cutoff
Twilio opened a minor incident at 02:57:20.918 UTC on 1 August for SMS delivery delays from a subset of its short codes to Claro subscribers in Colombia. At the fixed 03:17:59 cutoff, 20 minutes and 38.082 seconds had elapsed. There was no identified cause, recovery signal or resolution to report.
Anthropic closed a 47-minute Claude Sonnet 5 degradation with the service boundary still undisclosed
Anthropic marked a minor Claude Sonnet 5 performance incident resolved at 07:04:49.691 UTC on 31 July, 46 minutes and 41.623 seconds after it opened. That short chronology establishes a start and an end, but almost nothing about the customer experience between them: no symptom, access surface, geography, request count or technical cause was published.
Cloudflare’s us-east-1-aws path returned intermittent 5xx errors for 138 minutes
Cloudflare resolved a minor incident involving intermittent HTTP 5xx responses for customers using the label us-east-1-aws at 04:19:32.991 UTC on 31 July. A fix had entered monitoring 10 minutes and 51.068 seconds earlier. The operator disclosed the request outcome and chronology, but not the Cloudflare product, exact error codes, affected volume or root cause.
Twilio’s El Salvador incident delayed proof of SMS delivery, not necessarily the message
For 9 hours, 41 minutes and 49.987 seconds on 31 July, Twilio tracked delayed SMS delivery receipts for messages sent toward Telefónica subscribers in El Salvador. Its most important qualification was present from the first notice: message delivery might succeed while the receipt arrived late. Service was declared normal at 16:17:09.745 UTC, but the cause and delay distribution remained private.
A subset of Twilio numbers failed to deliver SMS to C+W LIME for more than 12 hours
Twilio closed a minor SMS incident affecting a subset of its phone numbers on the route to C+W LIME subscribers in the Cayman Islands at 06:18:08.210 UTC on 31 July. The case had opened the previous day and lasted 12 hours, 27 minutes and 40.379 seconds. The narrow route and origin subset are public; the size of that subset, cause and retry outcomes are not.
Twilio’s Celcom Malaysia SMS delay recovered without an identified public cause
Twilio resolved a 3-hour, 21-minute and 39.361-second incident affecting SMS delivery toward Celcom subscribers in Malaysia at 02:12:32.051 UTC on 31 July. The case moved from investigation to observed recovery and then resolution; it never displayed an identified state. No delay distribution, message population or root cause was published.
Twilio’s Claro Colombia voice incident changed its scope wording without measuring an expansion
Twilio resolved a minor voice-call incident affecting a subset of origin numbers toward Claro subscribers in Colombia at 02:09:34.991 UTC on 31 July. During the 5-hour, 41-minute and 10.935-second case, “Mobile Numbers” became “Phone Numbers” and one update omitted Claro. The record never says those edits represented a wider technical footprint.
Twilio was still monitoring Africell Angola SMS recovery after 25 hours at the cutoff
Twilio had observed recovery in SMS delivery toward Africell subscribers in Angola four times by the fixed 00:13:55 UTC cutoff on 1 August. It had not declared the minor incident resolved. The public case had been open for 25 hours, 42 minutes and 51.122 seconds, while cause, delay distribution, message scale and the remaining condition stayed undisclosed.
Fénix promises a 14 Tbps Venezuela–Curaçao ring, but the cable is still an announced project
Liberty Networks and CANTV announced Fénix on 31 July as a 378-kilometre submarine cable system between Camuri in Venezuela and Willemstad in Curaçao, with a stated minimum ring capacity of 14 Tbps. The design would connect Venezuela to a broader Caribbean hub. The release is evidence of a project launch, not of a cable already manufactured, laid, lit or carrying traffic.
Twilio’s Brazil voice incident reached monitoring after 18 hours without a public cause
Twilio said voice calls from its phone numbers to subscribers on multiple networks in Brazil were recovering on 31 July, but the incident was not resolved at the 20:19:07 UTC reporting cutoff. The first monitoring update came more than 18 hours after the case opened. Across that interval, the company acknowledged call failures and high post-dial delay, said it had identified the cause, and never disclosed what that cause was.
Twilio’s RCS and WhatsApp queues recovered after a 99-minute delay incident
Twilio recorded delays in both inbound and outbound delivery for RCS and WhatsApp on 31 July. The incident opened at 15:35:24.910 UTC, moved to monitoring at 16:44:39.656 and was declared resolved at 17:14:27.140. The chronology is clear; the scale and cause are not. Twilio disclosed neither a geography nor a message count, and “delay” does not establish loss, duplication or corruption.
Cloudflare’s London HTTP-error incident was still at the first diagnostic stage
Cloudflare opened a minor incident at 19:06:19.068 UTC on 31 July after detecting increased HTTP errors affecting a subset of customers. Fourteen minutes later, at the Wave 46 cutoff, the public record still said only that the company was investigating, analysing and mitigating the problem. The London label locates the incident, but the notice did not name a facility, route, product, error code, cause or customer denominator.
Hispasat’s Madrid teleport reaches WTA’s highest certification tier
Hispasat’s Arganda del Rey teleport has entered the World Teleport Association registry at Tier 4 Full, the programme’s highest level. The certification gives customers an independently inspected signal about operational discipline at a ground facility that links satellite control, uplink and terrestrial connectivity. It does not publish the audit scorecard, guarantee uninterrupted service or replace performance evidence from the facility itself.
Cloudflare applied a fix in Hamburg, but the incident was still under watch at cutoff
Cloudflare reported request errors or failures for customers whose traffic routed through its Hamburg location on 31 July. The company identified a problem at 15:20:19.991 UTC and said a fix was in place at 16:58:19.585. At the 17:49:33 Wave 45 cutoff, the incident remained in monitoring—not resolved. No affected product, traffic share, error rate or cause was disclosed.
Zoom’s Whiteboard and Tasks incident recovered, then returned
Zoom’s US-region status record shows two distinct degradation phases on 31 July. Users first lost the ability to list Zoom Whiteboards; Task creation was added to the affected scope fifteen minutes later. Both components entered monitoring at 12:45:35 UTC, and that was the state at this briefing’s 14:06:29 cutoff. Zoom then marked both services degraded again at 14:28:05, returned them to monitoring at 14:49:23 and closed the incident at 15:08:45. The record establishes a recurrence, but not its cause, reach or data impact.
Hamlet put 649 diesel engines through one public doorway
More than 100 people attended a North Carolina hearing on two air-permit applications tied to Amazon’s Energy Way Tech Campus in Hamlet. One application covers 592 Amazon emergency engines—588 critical units and four ancillary units. The other covers 57 Duke Energy engines intended as temporary bridge power before the campus receives its permanent grid connection. The hearing joined those systems for public scrutiny, but the state must decide on them separately. No final permit had been issued, and the engine count is not evidence that every unit is installed or operating.
Nebius has disclosed the carbon cost of expansion, but not its full footprint
Nebius reported 65,001 tonnes of market-based Scope 1 and 2 emissions for 2025, roughly 32 times the 2,036 tonnes disclosed a year earlier. The multiplication is striking, but its meaning depends on a changed operating base: the company says it expanded to seven active sites. The report makes the growth of operational emissions visible while leaving construction, hardware and much of the supply chain outside an uncompleted Scope 3 inventory.
Japan’s warning to LY Corporation is about a missing control path, not a data-sale finding
Japan’s Ministry of Internal Affairs and Communications issued written administrative guidance to LY Corporation after a LINE game-development partner transmitted MID identifiers and other specified user information to a third-party analytics service. The ministry says the path operated without LY Corporation’s approval and without the required user notice or confirmation opportunity. It counted about 8.03 million records, including about 7.52 million associated with users in Japan. The notice does not say that messages, passwords or payment details were exposed, or that the information was sold.
DOCOMO’s Ericsson choice moves the RAN bottleneck into compute
NTT DOCOMO has selected Ericsson RAN Compute as a baseband platform spanning 4G, 5G and 5G-Advanced. The decision matters because radio capacity is increasingly governed by processing, pooling and software lifecycle as well as antennas. Ericsson says the new generation can provide up to twice the capacity of its predecessor while using less than half the energy. Those are vendor comparisons, not measured results for DOCOMO’s network, and the public report does not disclose how many units will be deployed or when.
Veolia takes the operating risk in Ohio’s 350MW private power island
Veolia has been selected to operate and maintain a 350MW behind-the-meter microgrid for an undisclosed AI data-centre campus in Ohio. Gas engines and linear generators are expected to supply the campus, with a 430MWh battery supporting the system. The contract places Veolia at the point where a collection of power assets must become a continuously managed plant. It does not establish that 350MW of IT load is commissioned, that the availability target has been achieved or that an off-grid design removes fuel, emissions and equipment risks.
Cloudflare’s 70-minute incident separated its control plane from its edge
Cloudflare recorded a minor incident at 11:51:07 UTC on 31 July, initially warning that Dashboard and related-API requests might fail while Analytics was degraded. The event later expanded to Pages and Worker builds, entered monitoring at 12:43:57 and was marked resolved at 13:01:59. Cloudflare explicitly said cached-file delivery through its CDN and other Edge security features were unaffected. That separation matters: customers lost reliable access to management and build functions without evidence that ordinary cached traffic stopped flowing.
LG Uplus has funded a 200MW destination, not switched it on
LG’s additional KRW1.3tn commitment changes the financial scale of the Paju AI data-centre project. It does not change the physical state of the campus overnight. LG Uplus said in June that construction was about 20% complete and that a 200MW power supply had been secured; the new money therefore supports a defined route toward a very large facility rather than proving that 200MW of servers is already drawing power. The useful question is how capital, grid capacity, buildings and customers move through different clocks before a target completion in 2027.
ENGIE’s 7GW data-centre number is two pipelines wearing the same unit
ENGIE has moved a much larger share of its prospective data-centre work into an advanced stage: 4GW at the end of June, up from 0.8GW at the end of 2025. Together with 3GW at an early stage, that makes a 7GW project pipeline. Separately, the energy group also describes 7GW of power-purchase agreements with technology and data-centre companies. The matching number is a reporting hazard. One series tracks projects through development; the other tracks electricity contracts. Neither is 7GW of data centres already connected to a grid.
AirTrunk’s 400MW-plus SYD3 has a builder for its next phase, not a live meter
Bouygues Construction’s €1.2bn contract for the next phase of AirTrunk’s SYD3 campus is a large, bounded construction event. Awarded through Australian subsidiary A W Edwards, it assigns work and a delivery horizon toward mid-2028. It does not establish the final cost of the campus or 400MW of operating capacity. The announced scale belongs to the expanded design; the contract belongs to one phase; customer use will arrive only after construction, power, commissioning and acceptance have crossed their own thresholds.
Yondr is selling control of two Slough data centres without giving up the operating room
Yondr has agreed to sell a majority interest in a UK data-centre campus to GLIL Infrastructure, while retaining a minority stake and responsibility for operational management after completion. The asset consists of two fully operational hyperscale buildings in Slough serving one leading hyperscale customer. This is not a clean exit and it is not yet a closed sale. It separates economic control from day-to-day operation at an asset that Yondr had already taken into the securitisation market. The price, exact share split, customer, governance rights and treatment of £532m in term notes remain undisclosed.
OpenAI resolved a short chat incident but left its failure domain unnamed
OpenAI’s public incident record moved from identification at 09:04 UTC to mitigation at 09:06 and resolution at 09:28 on 31 July. The detailed update said some ChatGPT Business and Education users saw elevated errors when starting or continuing conversations. At this briefing’s fixed 09:23:13 window close, the incident was still in monitoring; the resolution arrived five minutes later and is included as an explicitly timed update. The visible status lifecycle lasted 24 minutes, but that is not necessarily the full customer-impact duration. OpenAI disclosed no failed component, cause, region, error rate, customer count or data-integrity effect.
CyrusOne appoints John Hatem as CEO
John Hatem has taken over as CyrusOne CEO, widening his authority from commercial and project delivery functions to company-wide strategy.
Veolia keeps €1bn data-centre and chip revenue target
CyrusOne’s appointment of John Hatem keeps an executive who linked customer commitments to procurement and construction in charge of its next expansion phase.
CityFibre shareholders propose £900m equity injection
CityFibre’s proposed funding ties network expansion and acquisitions to unresolved negotiations over its existing debt.
A reported $15bn loan would make Anthropic’s Texas compute a test of Google’s balance-sheet reach
The notable number in the reported financing talks for a data-centre campus in Hubbard, Texas, is not only the $15bn banks may lend. It is the obligation that could sit behind the debt. The Wall Street Journal reported that a Morgan Stanley-led consortium was discussing financing for Nexus Data Centers, while Google would guarantee billions of dollars of Anthropic’s lease and power commitments if the startup defaulted. Nothing in the public record shows that the financing has closed. Yet the proposed structure already reveals how AI infrastructure is moving beyond ordinary cloud procurement: lenders may be underwriting a chain that runs from a privately held model developer, through a project company and an onsite power plant, to the credit of a much larger technology group.
Paducah’s proposed AI campus joins a $100bn headline to a 17bn-dollar cleanup ledger
The U.S. Department of Energy has selected Brookfield Asset Management to develop and operate a proposed AI data-centre and energy complex at the government-owned Paducah Gaseous Diffusion Plant. AP reports a headline value of $100bn, with NextEra expected to build and own 2GW of gas generation, transmission upgrades and 2.6GW of battery storage for a 1.8GW AI campus. The site also carries a different long-duration number: a federal cleanup projected to continue to 2065 at about $17bn. Selection does not merge those obligations, approve the power agreement or identify the eventual computing customers. Kentucky’s lieutenant governor saying the governor’s office was not included in discussions exposes the missing governance layer between federal land, private capital and state-regulated electricity.
The Trans-Caspian cable has entered the seabed; it has not yet entered service
A specialised vessel has departed Baku and begun laying fibre toward Aktau, converting the Trans-Caspian project from stored equipment into a physical route across the Caspian Sea. Kazakhstan’s government says regulatory, organisational and customs obstacles were resolved, a legal equipment-export list was established, and roughly 600 tonnes of armoured cable arrived through Port Kuryk before transfer into the cable ship. CaspiLink, the Kazakhtelecom–AzerTelecom joint venture, expects the direct lay to take 15–20 days if weather is favourable and the full project to be commissioned by the end of 2026. Those are separate milestones. Seabed installation proves construction activity, not accepted landings, lit capacity or traffic.
Oregon cancelled the Salem land sale; Verrus says the data-centre proposal may still have another route
Gov. Tina Kotek has directed Oregon’s Department of Administrative Services to terminate the state land-sale contract tied to Verrus’s proposed data-centre campus at Mill Creek Corporate Center in southeast Salem. The proposal had been described at about $5.1bn, with three data-centre buildings and a substation on roughly 75 acres. Losing the state parcel is a material failure of site control. It is not identical to a city denial: Salem said the project remained at an early stage, no formal development application had been filed, and data centres were then an allowed use in the Employment Center zone. Verrus’s assertion that the action may not stop the project therefore cannot be dismissed, but it also does not identify a replacement parcel, contract or approved plan.
UNITEL has contained a cyber incident; Angola is still waiting for a complete recovery ledger
A national mobile network does not return with one switch. UNITEL says phased restoration of voice, data and internet began at 11:45 local time on 29 July after a cyberattack had disrupted those services across Angola. Its 30 July market notice described partial service in 13 provinces, SMS still unavailable and work continuing elsewhere. Hours later, Twilio was still reporting delivery delays and failures to UNITEL subscribers. The useful question is therefore not whether the attack is “under control”. It is which service works, where, at what quality, through which dependency, and for how long. Until UNITEL publishes that ledger, containment is an important milestone rather than proof of normal operation.
Southaven’s temporary turbine bridge finally has an exit calendar
SpaceXAI says an agreed order with Mississippi regulators now sets dates for removing all 69 temporary mobile gas turbines at its Southaven power site: work may begin in August 2026 and is scheduled to finish by July 2027. In parallel, a permitted permanent plant designed for 1.2GW and 41 turbines is under construction. The timetable converts an improvised source of data-centre power into a transition that can be tested month by month. It does not show that any turbine has yet left the site, that the replacement plant is complete, or that a change from temporary to permanent generation will reduce emissions or noise.