Scale & Strategy
together with
This is Scale & Strategy, the daily newsletter that’s the co-pilot for your flight through market turbulence.
Here’s what we’ve got for you today:
- How OpenAI Is Approaching AI's Cyber Red Line
- CoreWeave's Results Suggest the AI Infrastructure Boom Is Far From Over
How OpenAI Is Approaching AI's Cyber Red Line
As concerns about increasingly capable AI systems continue to mount, OpenAI is signaling that one of its upcoming models may have crossed an important internal threshold.
On Friday, the company disclosed that recent evaluations of Astra, an unreleased model, showed what it described as "significant advancements" in agentic coding and cybersecurity capabilities. Based on those results, OpenAI said it "cannot rule out" that Astra possesses what its Preparedness Framework classifies as critical cyber capabilities.
That distinction matters.
Under OpenAI's Preparedness Framework, a model reaches the "critical" cyber threshold if it can independently discover and develop zero-day exploits across hardened real-world systems or devise and execute novel cyberattack strategies against sophisticated targets. Previous OpenAI models, including GPT-5.6-Sol, were classified at the lower "high" capability level.
In response, OpenAI is implementing a series of additional safeguards around Astra, including:
- Isolated testing environments and stricter restrictions on network and tool access
- Suspending internal activities involving Astra that do not meet enhanced security requirements
- Universal monitoring for risky actions and signs of misalignment in agentic deployments
- Security guidance for third-party testing organizations
- Expanded collaboration with government agencies and AI safety groups to evaluate the model's capabilities
"We are sharing this because we believe it's important to be transparent with the public and the safety and security communities about this potential shift in capabilities," OpenAI said in its announcement.
The disclosure comes amid growing concern about the cybersecurity implications of increasingly autonomous AI systems. Recent incidents involving AI agents bypassing restrictions, manipulating environments, and demonstrating unexpected behaviors have intensified scrutiny of frontier labs and their safety practices.
OpenAI noted that Astra was not involved in the recent Hugging Face security incident, but the company's announcement reflects a broader reality: frontier AI models are becoming increasingly capable of performing tasks that were once the exclusive domain of highly skilled human operators.
The fact that OpenAI is slowing deployment and implementing additional controls is encouraging. But it should not be mistaken for an act of extraordinary restraint.
This is simply what responsible stewardship looks like when developing systems that may possess offensive cyber capabilities.
The real challenge is that every frontier lab is caught in the same tension. They are racing to build increasingly powerful systems while simultaneously trying to ensure those systems remain controllable. OpenAI, Anthropic, and others all want to lead the next era of AI. None of them wants to be remembered as the company that released the model that caused a major cybersecurity disaster.
For now, the safeguards appear to be holding.
The harder question is what happens as capabilities continue to improve. The industry's leading models are becoming more autonomous, more effective, and increasingly capable of operating in complex digital environments. Each successive generation pushes closer to thresholds that only a few years ago seemed theoretical.
OpenAI's Astra announcement is less important because of what the model can do today and more important because of what it signals about where AI is headed. Frontier labs are beginning to acknowledge that some models may soon possess capabilities that create meaningful cybersecurity risks. The challenge is no longer whether these systems will become powerful enough to warrant special controls. It's whether those controls can keep pace with the capabilities being created.
Sitting still isn’t on the agenda. With over 2,200 sessions, 70% are hands-on at AWS re:Invent, returning to Las Vegas, November 30 - December 4.
- Test a service against a realistic workload before you commit to it in production
- Ask a service engineer about an edge-case constraint
- Access AWS certification discounts, with exam prep built into the schedule
- Talk through implementations with 400+ partners on the expo floor
Learn more and sign up by August 25th to save $1,200.
CoreWeave's Results Suggest the AI Infrastructure Boom Is Far From Over
For a company that has become one of Wall Street's favorite AI bubble indicators, CoreWeave just delivered another data point that the demand side of the equation remains very real.
The cloud infrastructure provider reported $2.58 billion in quarterly revenue on Tuesday, marking its fifth consecutive quarter of record revenue and slightly beating analyst expectations. More importantly, the company revealed a sales backlog of $104 billion, nearly double what it reported last November.
Investors responded by sending the stock up roughly 13% in after-hours trading.
The backlog figure is what matters.
CoreWeave sits in a unique position within the AI ecosystem. The company buys GPUs from Nvidia, installs them in data centers, and rents computing capacity to customers including OpenAI and Microsoft. As a result, its order book serves as one of the clearest indicators of how much future AI infrastructure demand actually exists.
Despite widespread concerns about an AI spending bubble, CoreWeave continues to report demand that exceeds its ability to supply compute.
The company disclosed that it added another $25 billion of net new customer commitments early in the current quarter, on top of the $104 billion backlog already reported.
The debate surrounding AI infrastructure has gradually shifted over the past year.
Twelve months ago, investors were asking whether enough demand existed to justify the massive buildout of data centers, chips, and power infrastructure. Today, the bigger concern is whether companies can build capacity fast enough to satisfy the demand already in front of them.
Construction delays, component shortages, power constraints, and permitting challenges have become the primary bottlenecks.
Those concerns aren't theoretical for CoreWeave. Late last year, the stock fell nearly 50% in six weeks as investors worried about infrastructure delays and broader concerns that AI spending could eventually cool.
Yet the company's latest results suggest demand remains strong.
CoreWeave added 500 megawatts of computing capacity during the quarter and says it is bringing infrastructure online faster than originally projected.
"We're executing better than expected," said co-founder and Chief Development Officer Brannin McBee. "We're effectively sold out this year, and you can effectively extrapolate that to a lot of demand next year, too."
That doesn't mean the company is without risk.
CoreWeave remains deeply capital intensive and continues to post significant losses as it races to build infrastructure ahead of demand. The company reported a quarterly net loss of $626 million, bringing cumulative losses since its IPO to nearly $2.3 billion.
Investors continue to scrutinize its financing strategy, balance sheet, and path to profitability.
The challenge for CoreWeave is straightforward: secure enough GPUs, power, land, and financing to build capacity before competitors do, while ensuring customer demand remains durable enough to justify the investment.
So far, the numbers suggest demand isn't the problem.
The company recently secured a new $2.6 billion loan facility, while Nvidia announced a broader financing initiative that could provide up to $500 billion in debt financing for customers purchasing AI infrastructure. Both developments should help alleviate one of the industry's biggest constraints: access to capital.
The biggest question in AI infrastructure is no longer whether demand exists. CoreWeave's $104 billion backlog suggests it clearly does. The question is whether the industry can build enough data centers, secure enough power, and deploy enough compute fast enough to keep up. If CoreWeave is right that it's effectively sold out through this year and demand remains strong into next year, the AI infrastructure buildout may still be much closer to the beginning than the end.
Was this email forwarded to you?
That’s it for today and as always It would mean the world to us if you help us grow and share this newsletter with other operators.
Our mission is to help as many business operators as possible, and we would love for you to help us with that mission!
