Reuters · Jul 26
Nvidia in talks to backstop $250B in financing for OpenAI's 10-gigawatt Ohio data center
Nvidia is negotiating to guarantee roughly $250 billion in lease and debt financing so OpenAI can lease a 10 gigawatt data center that SoftBank's SB Energy unit is building in southern Ohio, a project expected to top $500 billion with chips included. Nvidia is separately discussing financing up to $350 billion of OpenAI chip purchases, and the first phase of about 800 megawatts is slated for 2028. The Wall Street Journal first reported the talks.
▸ The MSP Angle
Does it matter to my business that AI giants are financing each other?
It matters because circular financing concentrates risk. When the chip vendor guarantees the model vendor's debt, the health of your AI stack depends on a small set of intertwined balance sheets, and clients who lived through vendor consolidation in other markets will recognize the shape. The practical hedge is the same one that protects margin: build client services so the underlying model and compute supplier can change without a rebuild, and treat any single-vendor dependency as a risk you charge to manage.
Read at Reuters ↗
NicFab · Jul 24
The EU's Digital Omnibus is in force: high-risk AI rules pushed to December 2027, new deadlines land first
Regulation 2026/1744, the Digital Omnibus on AI, was published in the EU Official Journal on July 24 and enters into force July 27. It defers high-risk compliance under Annex III from August 2026 to December 2, 2027, adds new prohibitions on AI-generated intimate imagery effective December 2, 2026, hands the AI Office exclusive enforcement powers with fines up to 5% of worldwide turnover, and gives small businesses simplified documentation and capped penalties. Machine-readable content marking obligations for generative systems already on the market arrive December 2, 2026.
▸ The MSP Angle
Did the EU just delay the AI Act, and does that let my clients relax?
The high-risk rules slipped sixteen months, and the deadlines that remain got closer and more concrete. Content marking and the new prohibitions land this December, and enforcement just centralized under an office with real fining power. For clients selling into the EU, that is a calendar you can manage as a service: inventory which of their AI uses touch the new December obligations, then use the 2027 runway to get the high-risk work done unhurried. Relaxing is what their competitors will do.
Read at NicFab ↗
Anthropic · Jul 24
Anthropic ships Claude Opus 5 at half the price of its frontier model
Anthropic released Claude Opus 5 at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8, with a selectable effort setting to trade cost against capability and a fast mode at double the price for roughly 2.5 times the speed. Anthropic says it doubles Opus 4.8 on Frontier-Bench and lands within half a percent of Claude Fable 5 on a major coding benchmark at half the cost. It is the fourth Claude 5 family release in under two months.
▸ The MSP Angle
Model prices keep falling. When does that show up in my AI margins?
The day you adopt the cheaper model, and not before. Near-frontier output at mid-tier token prices only becomes margin if your delivery can switch models without a rebuild, which is an architecture decision you make once and collect on every release cycle. Four frontier releases in two months is the new cadence. Price your AI services against the capability the client receives, keep the model choice yours, and the vendors' price war becomes your expanding spread.
Read at Anthropic ↗
GTIA · Jul 22
GTIA: 97% of IT service providers have adopted AI, and 20% have the governance to profit from it
GTIA's new channel research, built on first quarter 2026 interviews with IT service providers, vendors, distributors, and end customers, finds 97% of organizations report AI adoption while only 20% have the governance frameworks, formal policies, and commercial strategies to turn it into durable business value. GTIA research vice president Carolyn April says AI has become a business imperative for channel firms but the path from adoption to business impact is rarely straightforward.
▸ The MSP Angle
Every MSP says they use AI now. What separates the ones making money on it?
Governance, priced and sold, separates them. The 77 point gap between adoption and operationalization is the whole opportunity: a written policy, controls on what data goes where, and a commercial model that bills for AI outcomes instead of absorbing them into existing plans. This is your own industry's data saying the field is wide open. Being in the 20% is a choice you can make this quarter, and it is also the exact service your clients need from you.
Read at GTIA ↗
Google · Jul 23
Google's ATLAS study: AI touches 68% of occupations but only 21% of their tasks
Google's AI and Economy ATLAS analyzed 14.65 million de-identified Gemini interactions across more than 150 countries, 140 languages, and 800 occupations. AI use now appears in 68% of detailed occupations, covering 88.4% of employed United States civilian workers, yet the median share of an occupation's tasks where AI shows up is just 21%, and full automation is the intent in under 10% of work conversations. Twenty-nine percent of occupations show no meaningful task coverage at all.
▸ The MSP Angle
Is AI actually doing people's jobs yet, or just hovering around them?
Hovering, by Google's own telemetry, and the gap is the business. AI reaches nearly every occupation but sits in a fifth of the work, mostly assisting rather than automating end to end. Getting from a fifth to real coverage takes workflow design, integration, and governance that employees will never do for themselves in a chat window. That is precisely the work a managed AI service sells. Bring the 68/21 numbers to your next client review and ask which of their tasks sit in the untouched 79%.
Read at Google ↗
The Hacker News · Jul 16
Agent Data Injection forges facts inside AI agents with up to 50% success, no prompt injection needed
Researchers at Seoul National University, the University of Illinois, and Largosoft disclosed Agent Data Injection, which plants punctuation like escaped quotes and dollar signs in attacker-controlled fields so language models misparse them as structure, letting attackers forge email sender names, button labels, or tool records without a classic prompt injection payload. The technique succeeded 31 to 43 percent of the time against major frontier models and up to 50 percent against purpose-built defenses, with browser and coding agents among the affected targets. The major vendors acknowledged the reports; no fixes were confirmed at publication.
▸ The MSP Angle
Can the AI agents my team uses be fooled by the data they read?
Yes, at the data layer, which is nastier than prompt injection because there is no suspicious instruction to filter. An agent that trusts a forged sender name will act on it, and the research shows the defenses that stop classic attacks barely slow this one. Treat agent-readable data the way you treat user input in web security: least privilege on what agents can act on, human confirmation on consequential actions, and logs that let you reconstruct what the agent believed. Architecture beats filtering here.
Read at The Hacker News ↗
ChannelPro · Jul 24
Druva ships AI Resilience as breach notices hit 471 million in six months
Druva launched AI Resilience, four capability sets for reversing AI-driven changes and protecting backup data from machine-speed risk, with chief executive Jaspreet Singh framing the test as whether you can reverse AI changes, breaks, or context loss. The same week's channel roundup carried the Identity Theft Resource Center's half-year count of 471.2 million breach victim notices, already past all of 2025's 297.5 million, and NCC Group data showing ransomware attacks up 3% quarter over quarter to 2,229.
▸ The MSP Angle
What happens when an AI agent breaks something and nobody can undo it?
That question now has a product category, which tells you the risk is real and billable. Agents change systems at machine speed, and recovery tooling is repositioning from ransomware rollback to AI rollback. Two moves: check whether your current backup stack can actually reverse an agent's changes with context intact, and put AI-caused damage into client recovery plans and tabletop exercises now. The breach notice numbers say the blast radius keeps growing either way.
Read at ChannelPro ↗
Upwork · Jul 14
Upwork: 38% of skilled knowledge workers now freelance, and AI work pays 34% more per hour
Upwork's Future Workforce Index, surveying 2,400 United States skilled knowledge workers, finds 38% now freelance, up from 28% a year earlier, with 58% of full-time employees considering the move. Freelancers doing AI work earn 34% more per hour, complex AI-augmented work saw 45% year over year earnings growth, and generative AI creative production grew 90% in contract volume while per-contract pay fell 13%.
▸ The MSP Angle
Is AI skill still a pay premium, or is that window closing?
Both, split by altitude. Commodity AI output is already deflating, with creative production volume up 90% and pay down 13%, while judgment-heavy AI work grew earnings 45%. For your hiring, the premium engineers now freelance, which cuts both ways: harder to keep on staff, easier to rent for a build. For your services, the same curve applies: anything a client could get from a cheap contractor will not hold a price, and the expertise-plus-context layer is where a managed offering defends its rate.
Read at Upwork ↗
Google Workspace Updates · Jul 16
Google Vids adds verified personal AI avatars with Gemini Omni
Google Vids now lets users cast a personal avatar, built from a likeness capture with account verification, in AI-generated video clips powered by Gemini Omni. Rollout began July 16 across Business and Enterprise tiers plus the AI Pro and Ultra add-ons, with admin console controls; personal avatars are English-only, restricted to adults, and unavailable in the European Economic Area, Switzerland, and the United Kingdom.
▸ The MSP Angle
Do I need a policy for AI avatars of my clients' employees?
You do now, because likeness generation just landed inside a mainstream business suite with an admin toggle you control. The governance questions are concrete: who may create an avatar, what may it be used for, what happens to the likeness capture when someone leaves, and is the feature enabled at all for regulated clients. Update acceptable-use policies and admin baselines before the first marketing team discovers the button, and note the regional restrictions if you manage EU tenants.
Read at Google Workspace Updates ↗
TechCrunch · Jul 26