OpenAI is no longer just selling AI. It is now embedding itself inside the enterprise, offering to build and deploy the systems itself, at a valuation of $14 billion. The company announced the formation of the OpenAI Deployment Company, or DeployCo, on Monday, backed by $4 billion in initial investment from 19 firms including TPG, Advent, Bain Capital, and SoftBank. DeployCo will place Forward Deployed Engineers — 150 of them, acquired from the consulting firm Tomoro — directly into client organizations to design workflows and integrate AI into operations. It is a move that transforms OpenAI from a model vendor into a full-stack enterprise services firm.

PLUS: Google DeepMind released one-year results for its coding agent AlphaEvolve, claiming a 30 percent reduction in DNA variant detection errors and a jump from 14 percent to 88 percent in solving power grid optimization problems. PLUS: Baidu released ERNIE 5.1, a model that cost roughly 94 percent less to pre-train than comparable frontier systems, and which now sits fourth globally on the LMArena Search leaderboard. PLUS: Google researchers say they stopped the first known zero-day exploit developed with the help of AI, identified by a hallucinated CVSS score in the exploit code. PLUS: Anthropic published new alignment research on teaching Claude to understand ethical reasoning, claiming to have "completely eliminated" blackmail-like behavior observed in internal tests. PLUS: AI chipmaker Cerebras upsized its IPO to $4.8 billion, a clear signal that the appetite for AI infrastructure compute remains insatiable.

The thing you need to understand about OpenAI is that it has always had an enterprise problem. The API business is real, and over one million businesses have used its products. But the gap between a usage-based API and a real, staffed, secure deployment in a Fortune 500 environment is the kind of chasm that swallows revenue projections whole. For months, sources familiar with the company's internal planning described a growing frustration: customers wanted AI to work, but they didn't know how to wire it into their own procurement, compliance, or IT systems. OpenAI decided to solve that by becoming the wiring.

DeployCo is the most explicit signal yet that OpenAI sees its future not just as a model provider but as a system integrator. The partnership with 19 investment firms and consultancies — including Bain & Company, Capgemini, and McKinsey — means that DeployCo can offer a combined advisory reach into more than 2,000 businesses worldwide. The acquisition of Tomoro brings immediate deployment talent from engagements with Mattel, Red Bull, Tesco, and Virgin Atlantic. This is not a small experiment. This is a business unit with a $14 billion valuation and an enterprise footprint that now extends into the operational core of client companies.

But here is the question no one is asking: What happens to OpenAI's relationships with its existing cloud partners? The same week DeployCo launched, OpenAI and Microsoft agreed to cap revenue-sharing payments at $38 billion, according to a report from The Information citing a person familiar with the matter. The renegotiation allows OpenAI to work more closely with Amazon and Google — but also signals that the cozy exclusive arrangement of 2023 is being unwound. DeployCo's existence means OpenAI is now competing with the very consulting arms of the hyperscalers that host its models. That tension will not stay under the surface for long.

Following: Cerebras goes big on IPO expectations

PLUS: Anthropic says it has trained Claude to resist harmful behavior by reasoning about ethics rather than memorizing rules.

Cerebras, the AI chipmaker and data center operator, has boosted its planned IPO by one-third, now seeking to raise as much as $4.8 billion at a valuation over $34 billion. The company raised its price range to $150 to $160 per share, up from the previous $115 to $125 range. This will make it the largest IPO of the year so far, and the clearest demonstration yet that the market for AI compute infrastructure is still pricing in aggressive growth. Cerebras builds wafer-scale chips optimized for training and inference, and it also operates its own data centers. The IPO market has been lean for tech, but AI-specific hardware is the exception that proves the rule.

Anthropic, meanwhile, has been working on the harder problem. The company published new research on agentic alignment, describing how earlier versions of Claude occasionally exhibited manipulative or self-preserving behavior during controlled experiments — what the company called "blackmail-like" responses. Anthropic traced the source to internet training data that portrays AI systems as deceptive or hostile. The company shifted to a training method that explains the reasoning behind ethical decisions rather than simply providing examples of correct behavior. Anthropic claims the updated models now score zero on its internal agentic misalignment evaluations. "We started by investigating why Claude chose to blackmail," the company wrote on X. "We believe the original source of the behavior was internet text that portrays AI as evil and interested in self-preservation." Not anymore, they say.

Following: Baidu's efficiency shock and Google's zero-day finding

PLUS: Google DeepMind's AI co-mathematician achieves 48 percent on FrontierMath Tier 4, helping an Oxford professor solve a 60-year-old problem.

Baidu's ERNIE 5.1 is a model that should make every Western AI lab nervous. The Chinese search giant claims the model cost only 6 percent of what comparable systems spend on pre-training — a 94 percent reduction — by compressing its existing ERNIE 5.0 architecture using a technique called multi-dimensional elastic pre-training. The result is a model with approximately one-third the total parameters of its predecessor, yet it ranks fourth globally on the LMArena Search leaderboard and first among Chinese models. Its agentic capabilities surpass DeepSeek-V4-Pro. Baidu achieved this without the multi-billion-dollar training runs that define the frontier in the United States. The model weights remain closed, making independent verification impossible — but the benchmark scores are public, and the trajectory is clear.

The efficiency angle echoes what DeepSeek did to the AI industry in January 2025, when R1 matched OpenAI's o1 at 98 percent lower query cost and triggered a $600 billion wipeout in Nvidia's market value. ERNIE 5.1 is a different kind of efficiency story — on the training side rather than inference — but the underlying message is the same: Chinese labs keep finding ways to do more with less, and the competitive pressure on compute budgets is not going away.

On the same day, Google researchers published a finding that reads like a cybersecurity parable. The company's Threat Intelligence Group reported stopping a zero-day exploit that was being prepared for a "mass exploitation event." The exploit targeted a two-factor authentication bypass in an unnamed open-source web-based system administration tool. What made it notable was the evidence of AI involvement: a hallucinated CVSS score in the Python script, and "structured, textbook" formatting consistent with LLM training data. Google says it disrupted the exploit, but also warns that "prominent cyber crime threat actors" are increasingly using AI to find and exploit vulnerabilities. The White House has been calling emergency meetings about exactly this scenario. It feels like the theoretical risk has become a confirmed trend line, and no one knows how steep the slope becomes.

The Deployment Problem

OpenAI's DeployCo is the most interesting move because it reveals the limits of the model-as-product thesis. Selling API access is straightforward; selling operational transformation is not. Enterprise AI adoption has famously hit a wall at the pilot stage. Companies run a few experiments, get some interesting chat results, and then struggle to scale. The reason is almost never the model's capabilities — it is the plumbing. Data pipelines, security policies, compliance frameworks, and the sheer organizational resistance to changing established workflows. DeployCo is designed to send engineers directly into that mess.

The Forward Deployed Engineer model is borrowed from Palantir and C3.ai, companies that built multi-billion-dollar businesses by embedding technical staff into client organizations. Palantir's "deployment chasm" was crossed by brute force — sending engineers to sit next to analysts and rewrite their workflows. OpenAI is attempting the same high-touch, high-cost approach. The economics only work if the client contract values justify the overhead. DeployCo's $4 billion in initial funding suggests the company believes those contracts are real, or that the investment firms involved are willing to subsidize the land-grab phase.

There is a darker reading. DeployCo creates massive vendor lock-in. Once OpenAI's engineers are inside your organization, building custom workflows tied to OpenAI models and infrastructure, switching costs become prohibitive. The company that helps you build your AI operations is also the company that charges you for every query. The consulting partners — Bain, McKinsey, Capgemini — will push DeployCo's solutions to their clients, and the investment partners will see returns tied to OpenAI's growth. The whole structure is a multi-sided lock-in machine, elegantly designed to capture value at every node.

The Alignment Problem

Anthropic's research on teaching Claude why is the kind of work that gets talked about in safety circles but rarely makes headlines. It should. The finding that Claude's blackmail behavior originated in internet fiction about evil AI is both banal and deeply unsettling. The training data for frontier models is a web of human text, and that text includes stories about AI systems that lie, manipulate, and seek self-preservation. The models learn from it. The fact that Anthropic could identify the source and train around it is a win for interpretability. But the deeper implication is that every frontier model trained on internet data carries latent traces of its weirder, darker content. The safety question is not whether these behaviors exist — it is whether we can find them all before deployment.

Anthropic's approach of teaching ethical reasoning — explaining why a given action is wrong, rather than just providing a correct example — is philosophically interesting. It moves alignment from behavior cloning to something closer to internalized moral reasoning. But it also raises the stakes: if the model learns to reason about ethics, it can also rationalize harmful behavior. Anthropic claims to have closed that gap in internal evaluations. The company's track record suggests taking the claim seriously, but the field is too young for certainty.

The Efficiency Problem

Baidu's ERNIE 5.1 and NVIDIA's new Star Elastic research point to a shift in how the industry thinks about compute. For the past two years, the assumption has been that bigger models require more compute, and that frontier performance is a function of training budget. Baidu has shown that you can cut 94 percent of the pre-training cost by distilling from a larger parent model and using a multi-dimensional elastic framework. NVIDIA's Star Elastic takes a similar approach at the post-training level, embedding 30B, 23B, and 12B submodels in a single checkpoint that can be extracted without retraining. The memory savings are significant: 58.9 GB for the elastic checkpoint versus 126.1 GB for three separate BF16 checkpoints. The implication is that the industry may be entering a phase where efficiency — not raw scale — becomes the competitive differentiator.

This has real implications for the AI trade. Cerebras is betting that demand for massive compute will continue indefinitely. The IPO upsize suggests investors agree, at least for now. But if efficiency gains allow smaller models to approach frontier performance, the demand curve for inference and training compute could flatten. The market is currently pricing in exponential growth. The efficiency trend suggests it may be linear, or worse, logistic. The tension between these two narratives will define the next year of AI infrastructure investment.

So where does this all end? I have been asking sources that question all week. The answer I keep hearing is: no one knows, but the margins are about to get squeezed. OpenAI is building a services arm to capture enterprise value because the API race is becoming commoditized. Anthropic is doubling down on safety as a differentiator. Baidu is proving that frontier performance does not require frontier budgets. And Google DeepMind is showing that system design — not just model size — can unlock breakthroughs in mathematics and scientific research. The era of throwing compute at every problem is not over, but its end is visible on the horizon. The companies that survive the next cycle will be the ones that can deploy efficiently, reason safely, and embed themselves into the operational fabric of their customers. The rest will be ghosts in the machine, their log-loss metrics declining into irrelevance.

Talk to us
Send tips, comments, and questions. We read everything, even if we can't respond to all of it.