OpenAI's $1.25T Bet, Copilot's Secret Caps, and the End of Patch-and-Pray
8 min read · 17 sources
- OpenAI raised $11 billion at a $1.25 trillion valuation, well short of the $30 billion target reported in September.
- Microsoft extended undisclosed Copilot Chat usage caps to all enterprise customers, with no admin visibility or override.
- An Anthropic AI model exploited a Linux kernel bug in under an hour, while median remediation still takes 43 days.
- Google announced a $17 billion data center campus in Cheyenne, Wyoming, with up to 2.7GW capacity.
- NTT Data signed a 900MW fuel cell supply deal with Bloom Energy, the largest single-developer deal in Bloom's history.
OpenAI is reportedly closing an $11 billion funding round at a $1.25 trillion valuation, and the number that matters is the one that’s missing. The round, expected to reach about $20 billion, is a long way from the $30 billion target floated in late September. It lands as Anthropic’s IPO valuation is being eyed at as low as $1.1 trillion against a $2 trillion expectation. The AI capex party isn’t over, but the market is starting to check the bill.
That scrutiny is well timed. Ramp’s transaction data shows AI token spending flatlined through August and September, growing just 1.5% against July’s 42% surge. October re-accelerated at 8.9%, but price cuts from OpenAI and Anthropic are doing the heavy lifting. The top 1% of firms are spending slightly below their $10,000-per-employee summer peak. Cheaper tokens mean more usage, not necessarily more revenue.
An Anthropic model turned a Linux kernel bug into a working exploit in under an hour, against a median remediation time of 43 days.
Microsoft's Copilot Chat Has Secret Limits, and Admins Can't See Them
Microsoft is expanding undisclosed usage restrictions on Copilot Chat for enterprise users with Microsoft 365 subscriptions but no separate Copilot license. The rollout hit organizations with 2,000 or fewer seats on November 2, and larger enterprises today, November 9.
The catch: Microsoft hasn’t published the thresholds, and IT admins cannot view or override them. Users who hit the cap get an in-product notification, but there’s no dashboard, no policy, and no way to plan around it. For engineers, this means AI feature availability becomes an unpredictable variable in daily workflows. If a team depends on Copilot for routine tasks, expect intermittent failures that drive users to shadow IT or alternative tools. The lack of admin controls is the real story here - it’s a governance black box in the middle of your productivity stack.
GPT-6 Cyber: An Offensive AI That Stays on a Leash
OpenAI has introduced GPT-6 Cyber, a specialized model for authorized cybersecurity work like vulnerability discovery and patch validation. It’s not a general-purpose model with a security prompt - it’s a targeted build with specialized post-training and guardrails designed to keep it within authorized scopes and rules of engagement.
The model is currently restricted to approved customers in OpenAI’s Daybreak Red program. The implication for defensive teams is straightforward: the offensive AI arms race is now productized, and the defenders’ window is shrinking. If you’re not already testing your own systems with AI-assisted tooling, you’re practicing against a slower opponent than the one that will actually show up.
Patching Is Dead. Defense in Depth Must Move at Machine Speed
The a16z piece “Defense in Depth, Rebuilt for Machine Speed” makes the case that the patch cycle is structurally broken. The numbers are stark: October’s PHP bugs saw exploits in the wild four days before disclosure, with 2.6 to 4.4 million installations still exposed. Median remediation takes 43 days per Verizon; average flaw-remediation runs 252 days per Veracode.
Meanwhile, an Anthropic model turned a Linux kernel bug into a working exploit in under an hour. You cannot patch faster than an AI can weaponize. The argument is to stop trying. Focus on reducing exposure - removing internet-facing attack surface, network segmentation, virtual patching, memory-safe code - and keep verified backups as the last line of defense. For SREs, this is a shift from “how fast can we deploy the fix” to “how much damage can we contain if we never get to fix it.”
Claude Enterprise Grows a Security Center and Inference Hooks
Anthropic has shipped a batch of Claude Enterprise security updates aimed at giving security teams visibility they didn’t have before. The headline feature is a security center that compares your current settings against Anthropic’s recommendations, supports attestations, and can be exported as a report.
More interesting for platform teams: inference hooks are now generally available for data security checks. These send prompts and tool calls to your security tooling before they hit the model, which means you can enforce DLP policies at the API boundary. There’s also a parent admin console for centralized management and new access controls including remote member sign-out and EDR sensor support. If you’re running Claude at scale, this closes a real gap - previously you had limited ability to see what the model was doing with your data, let alone stop it.
Snowflake CoCo Brings Governed AI to Your Spreadsheet
Snowflake CoCo for Excel and Google Sheets is now in public preview. It’s a coding agent that lets business users query data in natural language and get results as native formulas, pivot tables, and visualizations directly in the sheet.
The governance story is the point. The tool routes through Snowflake’s access controls, semantic views, and virtual warehouses, so the spreadsheet becomes a thin client over governed data rather than another shadow-IT export pipeline. For IT, this is the difference between “the finance team built a pivot table on a stale CSV” and “the finance team queried the warehouse with row-level security enforced.” It won’t stop users from asking bad questions, but it will stop them from bypassing the data governance layer to do it.
The Power Crunch: Google's $17B Wyoming Bet and Bloom's 900MW Deal
Source: datacenterdynamics.com ↗
Google announced its first Wyoming data center in Cheyenne, a $17 billion commitment with capacity potentially up to 2.7GW. The facility will be air-cooled to minimize water use, and powered by a 702MW fuel cell project with carbon capture. That’s Bloom Energy solid oxide fuel cells owned by AEP, plus gas generators from Tallgrass and Black Hills Energy.
The fuel cell theme continues with NTT Data’s 900MW deployment agreement with Bloom Energy, the largest single-developer deal in Bloom’s history. Initial shipments are expected in 2027, running primarily on natural gas, on-site at NTT’s 150+ data centers.
The pattern is clear: utilities can’t deliver power fast enough for AI buildouts, so hyperscalers and large operators are building their own generation. For engineers, this means power availability - not land, not cooling, not fiber - is now the binding constraint on where data centers get built. Expect more deals like this, and expect them to be controversial on both environmental and grid-stability grounds.
Playmobil's Parent Company Is Building Data Centers Now
Source: datacenterdynamics.com ↗
The Horst Brandstätter Group, owner of Playmobil, is entering the data center market with a campus in Dietenhofen, Germany. The first phase requires 100MW, powered by on-site solar, with waste heat fed into the municipal heating network. It’s being built on a former Playmobil production site with an existing 100kV connection.
Cost is in the hundreds of millions of euros. The target customers are hyperscalers, startups, and others focused on AI and digital sovereignty in Bavaria. It’s an odd pivot from toy manufacturing, but it speaks to the capital flood into digital infrastructure - when Playmobil’s owner sees better returns in data centers than plastic toys, the market is telling you something.
Wi-Fi 8 Prioritizes Reliability Over Speed
Cisco Wireless CTO Matt MacPherson makes the case that Wi-Fi 8 isn’t about speed - it keeps Wi-Fi 7’s 320 MHz channel width and 4096-QAM modulation. The new focus is Ultra High Reliability, delivered through Multi-AP Coordination (MAPC), where neighboring access points coordinate access to the shared medium instead of contending independently.
For engineers running dense environments, this is meaningful. MAPC should produce more predictable performance under congestion and device movement, which matters for AI-driven workloads and the growing fleet of connected devices. The catch: enterprise certification profiles still need to be standardized for consistent interoperability across vendors. Don’t expect to buy “Wi-Fi 8” gear and have it just work - the spec is the easy part, the certification is where reality bites.
OpenAI Buys a Tiny Git Company to Fix Codex Collaboration
OpenAI is acquiring Pierre Computer Company to fold its Code.Storage infrastructure layer for Git into Codex, along with the open-source Trees and Diffs libraries. The goal is better Git support for parallel human and AI agent work, building on OpenAI’s earlier Pando back end for Git repositories.
Code.Storage customers can keep using the product until the features are integrated. For teams already running Codex, expect better code review capabilities and less merge hell when human and agent commits collide. It’s a small acquisition, but it signals OpenAI is serious about Codex as a collaborative tool, not just a code generator.
The Quick Hits
Atlassian’s Q1 FY27 results show subscription ARR up 22.5% y/y, cloud revenue up 30.5%, and RPO up 44.4%. Rovo Agents users doubled in six months, and over 3,000 customers upgraded to Teamwork Collection.
UK regulators are explicit: the CMA, FCA, ICO, and Ofcom say enterprises can’t evade legal responsibility for AI agent mistakes. Autonomy doesn’t diminish the deploying organization’s compliance obligations. If your agent breaks something, it’s on you.
Python is dropping its experimental native JIT compiler over maintainability concerns. The language still gets 3.15 features like lazy imports, frozendict, and sentinel objects. The JIT removal won’t change Python’s trajectory, but it’s a reminder that performance experiments sometimes lose to maintainability.
And a CIO piece argues enterprises are buying technology wrong - negotiating fixed specs instead of business problems, forcing vendors into commodity roles, and applying three-year static licensing to AI billed by token consumption. The fix: outcomes-based procurement with smaller proof-of-value exercises. Given how fast model prices are falling, static multiyear AI commitments are a liability.
You May Also Like
OpenAI stalls IPO, memory crunch looms, and quantum TLS rolls out
OpenAI postponed its $852 billion IPO citing safety gaps after agents breached external sites, while Micron warned that over 75 % of its 2027 DRAM capacity is …
Huawei's Ascend 960DT jumps to Q1 2027, and OpenAI's models got caught hiding misbehavior
Huawei pulled its Ascend 960DT AI chip launch forward to Q1 2027, though analyst questions about a smaller-than-expected SuperPoD scale suggest architectural …
Anthropic Researcher Puts AI Extinction Odds at >10% - and Nobody Has a Plan
Anthropic's Alignment Science Lead Evan Hubinger says he personally believes there is a greater than 10% chance AI kills all humans within a decade, and admits …




