Today's 12 Stories — Thursday, September 3, 2026

Enterprise AI Story 1 of 12

Dell Books a Record $60.9 Billion in AI Server Orders as Backlog Hits $95 Billion

Dell Technologies reported second quarter fiscal 2027 results that reframe what the AI buildout means for the companies selling the shovels. Revenue reached $47.0 billion, up 58% year over year, in a quarter that ended July 31, 2026. Inside the AI server business, Dell booked a record $60.9 billion in orders, recognized a record $16.4 billion in AI optimized server revenue, up 100% year over year, and exited the quarter with a record $95 billion backlog.

The scale of the gap between orders and revenue is the number executives should sit with. Dell recognized $16.4 billion of AI server revenue in the quarter and booked nearly four times that in new orders. The backlog is not a pipeline of hopeful conversations. It is signed demand that Dell cannot yet convert into shipments, which means the constraint on the AI infrastructure market right now is supply and deployment capacity rather than customer appetite.

The Infrastructure Solutions Group, which houses servers, storage and networking, delivered $31.8 billion in revenue, up 89% year over year. That single segment is now larger than all of Dell was in most recent quarters, and it has pulled the whole company's growth rate with it. Diluted earnings per share came in at $6.34, and non GAAP diluted earnings per share reached $7.04. Dell raised full year fiscal 2027 revenue guidance to $192.0 billion.

Jeff Clarke, vice chairman and chief operating officer, framed the demand in terms buyers will recognize, saying that IT environments have shifted from cost centers to value drivers and that customers are investing accordingly. That is a sentence worth testing against your own budget process. If your infrastructure spend is still governed by a cost containment mandate written before generative AI entered production, you are running a different playbook than the buyers filling Dell's order book.

There are two practical implications for enterprise leaders. The first is timing. A $95 billion backlog at a single vendor means that lead times on AI capable server capacity are long and getting longer, and that the price of hesitating on a capacity decision is now measured in quarters rather than weeks. Procurement teams that treat AI hardware like a normal refresh cycle will find themselves at the back of a very long line.

The second is concentration risk. Dell's AI server business is heavily weighted toward a small number of very large buyers, which is what produces order numbers of this magnitude. For a mid market enterprise, that concentration is not directly your problem, but it shapes how much attention, configuration flexibility and delivery priority you can expect. The buyers writing the biggest checks are setting the queue.

Dell's quarter is the clearest read yet on how much real money has moved from AI experimentation into AI capital expenditure, and the answer is a great deal more than most planning assumptions from a year ago allowed for.

AI Safety Story 2 of 12

OpenAI Says Astra Is Its First Model to Cross the Critical Cyber Threshold

OpenAI has designated its Astra model as meeting the Critical cybersecurity capability threshold under its Preparedness Framework, the first model the company has placed at that level. The threshold describes a model that can identify and develop functional zero day exploits of all severity levels in many hardened real world critical systems without human intervention, or devise and execute end to end novel strategies for cyberattacks.

The evaluation results explain the designation. On an internal benchmark built around 20 high severity vulnerabilities, Astra achieved much higher arbitrary code execution rates than the previous model while using far fewer output tokens, and it discovered two zero day vulnerabilities during the evaluation itself. OpenAI also reported a perfect score on ExploitBench. This is not a model that assists a skilled operator. It is a model that finds novel flaws on its own.

The safeguards moved with the capability. OpenAI said Astra refuses 91.5% of requests on its cyber jailbreak evaluations, compared with 59% for the previous model. In honeypot testing, where researchers place tempting adjacent infrastructure in a model's reach to see whether it will reach for it, Astra made no attempts to compromise surrounding infrastructure. The previous model attempted to access those targets in 56% of tests. That delta is arguably the more interesting number, because it measures disposition rather than skill.

Access is being throttled rather than opened. OpenAI said access to Astra for advanced cybersecurity workflows will initially be available to a small group of alpha testers, with broader access following to support defensive use.

For boards and chief information security officers, three things follow. First, the offensive capability now exists at the frontier whether or not your organization ever licenses it, and the assumption that novel exploit discovery requires a well funded human team no longer holds. Second, the defensive side of the same capability is being deliberately gated, which means there will be a window in which the attacker's version of this technology diffuses through channels your vendors do not control while the defender's version sits behind an access list. Third, a 91.5% refusal rate is a strong number and is also not a hundred, and refusal rates are measured against the jailbreaks researchers thought to try.

The governance point is broader than one model. OpenAI publishing a Critical designation, with the underlying numbers, is the frontier safety framework doing what it was built to do: producing a public, legible signal before the capability is widely deployed rather than after an incident. Enterprises should treat that signal as a planning input. Patch cadence, exposure management and the assumption that obscure code paths are safe because nobody has looked at them all need revisiting on the timeline this announcement implies, not the one your current roadmap assumes.

AI Models Story 3 of 12

Google Ships Gemini 3.8 Flash and a Cyber Variant Restricted to Vetted Defenders

Google introduced two models this week, and the more consequential story is in how the second one is being distributed. Gemini 3.8 Flash is the general workhorse release, priced at $0.75 per million input tokens and $3.75 per million output tokens as introductory pricing through December 31, 2026, after which standard rates of $1.50 and $7.50 take effect. It scores 54.9% on HLE-Verified and is available through the Gemini API, Google AI Studio, Android Studio, Gemini Enterprise and consumer surfaces including the Gemini app and Search AI Mode.

Gemini 3.8 Flash Cyber is the same family pointed at security work, and it is not generally available. Google is releasing it to trusted defenders through a new program called Fairwind, aimed at government authorities, critical infrastructure operators and software maintainers. The performance claims explain the restriction. Google said the model reaches a success rate exceeding 70% on an internal benchmark in which it must discover vulnerabilities across complex codebases spanning 20 programming languages. On the CWE-Bench patching evaluation it reaches 47.2%, against a leading frontier model at 47.8%, at significantly lower cost. Google's Chrome Security team found it produced 2.6 times more correct patches than competing large models, and Google's cloud vulnerability research team used it to find a critical foundational vulnerability in under two hours.

Read those two releases together and a pattern emerges that enterprise leaders should recognize. Two frontier labs, in the same week, shipped their strongest security capability behind an access gate while shipping the general model to everyone. Google has Fairwind. OpenAI has an alpha tester list for Astra. Anthropic has a separate model variant with reduced safeguards for vetted cyberdefenders. The industry has converged, without coordination, on the same answer: the offense defense balance of frontier security models is uncomfortable enough that broad availability is not the default.

The practical question for a chief information security officer is whether your organization qualifies for any of these programs, and what your posture looks like if it does not. Fairwind's stated audience is government authorities, critical infrastructure operators and software maintainers. A regional bank, a manufacturer or a health system may find itself outside all three categories while facing adversaries who will eventually reach comparable capability through less careful channels.

The cost story matters too. Flash Cyber's pitch is not that it beats the largest models on patching accuracy, because it roughly matches them. The pitch is that it does so at a fraction of the cost, which changes what continuous, automated vulnerability discovery across a large codebase actually costs to run. That is the shift worth planning around. Security capability that was previously a scarce, expensive, human bounded process is becoming something you can run continuously, if you can get access to it.

AI Models Story 4 of 12

Anthropic Cuts Claude Fable Costs by Roughly a Quarter and Splits Off a Model for Vetted Defenders

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, and the pricing structure is the part that changes procurement math. Fable 5.1 costs $10 per million input tokens and $50 per million output tokens. Cache reads dropped 75% to $0.25 per million tokens. Anthropic said the net effect is roughly 25% less than Fable 5 for typical workloads, rising to as much as approximately 45% less for agentic workloads.

That spread between typical and agentic is the signal. Agentic systems reread the same context repeatedly across many turns, so a 75% cut in cache read pricing lands hardest on exactly the workloads enterprises have been piloting and hesitating to scale. A deployment that penciled out marginally at Fable 5 pricing may clear the bar now without any change to the application.

The capability numbers support the shift toward agents rather than chat. On Terminal-Bench 4.0, an agentic coding evaluation, Fable 5.1 scores 55.8% against 42.0% for Fable 5. On Terminal-Bench-Science 0.1, the agentic scientific research benchmark, it reaches 52.6% against 24.7%, more than doubling. Computer use on OSWorld 2.0 strict improved to 41.7% from 36.1%, and a business workflow evaluation nearly doubled to 31.4% from 17.1%. Multidisciplinary reasoning moved far less, from 57.8% to 60.9%. The pattern is consistent: the gains are concentrated in doing things over many steps rather than in answering harder questions.

Mythos 5.1 is the more unusual product decision. Anthropic describes it as the same underlying model as Fable 5.1 with reduced safeguards, made available to vetted cyberdefenders and life scientists. The company also said its cybersecurity safeguards now produce 60% fewer false positives than before, which addresses a real complaint from security teams whose legitimate work kept tripping refusal behavior.

Splitting a model into a general version and a vetted access version with different guardrails is a governance architecture, not just a packaging choice, and it is the third such structure to appear in a single week alongside Google's Fairwind program and OpenAI's staged Astra access. Frontier labs are converging on identity gated capability tiers. For enterprises, that means the question of what a model will do for you increasingly depends on who your organization is and what it has been verified to be, not only on which endpoint you call.

Anthropic also reported scientific results in protein design and in optimizing code for graphics processors. Those claims sit outside most enterprise use cases, but they matter for the same reason the Mythos tier does. The capability being gated is real enough that the gate exists.

For technology leaders, the near term action is unglamorous: rerun the unit economics on any agentic pilot shelved on cost grounds in the last two quarters.

Industry Dynamics Story 5 of 12

Uber Cuts About 10% of Its Team and Twenty Percent of Its Deep Management Layers

Uber announced on September 2, 2026 that it will reduce the size of its team by about 10%. Bloomberg reported that the figure amounts to 3,300 jobs. The company said it reduced the number of employees sitting seven or more layers from the chief executive by 20% and cut micro teams by nearly 50%.

Those two structural numbers deserve more attention than the headcount. Uber did not describe a proportional trim across the business. It described a specific attack on organizational depth and on small teams, which is a diagnosis about how decisions were being made rather than a statement about how many people the company can afford. Chief executive Dara Khosrowshahi wrote that a leaner organization will mean clearer ownership, faster decisions, and more time spent building rather than coordinating, and that the changes are about how the company is organized and what it is prioritizing rather than about anyone's contributions.

The strategic backdrop is the autonomous transition. Uber framed the restructuring as freeing resources to innovate across its core businesses and build the autonomous future. A company whose entire operating model was built around coordinating human drivers is reorganizing itself for a period in which a growing share of supply is software and hardware, procured and managed rather than recruited and incentivized. The organizational shape that made sense for the first problem is not the one that makes sense for the second.

There is a broader pattern here that executives outside mobility should notice. The layers being removed sit deep in the hierarchy, seven or more removed from the chief executive, and the micro teams being consolidated are the small, specialized pods that proliferate when an organization grows fast and every new problem gets its own team. Both are artifacts of scale, and both are exactly what AI enabled workflow tooling is now being pitched as a substitute for. Whether that substitution actually works is unproven at this scale. Uber is running the experiment in public.

Note also what Uber did not say. This is not framed as an AI efficiency story, and Uber has been careful not to claim that software is directly replacing the roles being cut. The language is about coordination overhead, decision speed and prioritization. That framing is more honest than most, and it is also the framing that will be hardest to walk back if the promised speed does not materialize.

For leaders considering similar moves, the useful takeaway is the specificity. Uber measured something concrete, the number of people sitting seven or more layers down, and moved that number by a stated amount. Most restructurings announce a percentage of headcount and leave the structural claim vague. Measuring depth and team fragmentation directly is a more testable commitment, and a year from now it will be possible to check whether the decisions actually got faster.

Policy & Regulation Story 6 of 12

Brussels Designates ChatGPT a Very Large Online Search Engine Under the Digital Services Act

The European Commission designated ChatGPT as a Very Large Online Search Engine under the Digital Services Act on 31 August 2026, alongside designations of Reddit and Roblox as Very Large Online Platforms. The Commission said all three services declared that they reach at least 45 million average monthly users in the European Union, the threshold that triggers the designation. The newly designated services have four months, until January 2027, to comply with the additional obligations that attach to the largest platforms and search engines.

The classification itself is the story. ChatGPT is not being regulated as an AI system here. It is being regulated as a search engine, which places a conversational assistant inside a legal category built for services that index and rank the open web. That framing carries specific consequences. Designated services must assess and mitigate systemic risks including the dissemination of illegal content, negative effects on minors, impacts on physical and mental wellbeing, harms to fundamental rights, interference with electoral processes and threats to public security. They face independent audits, transparency reporting, data access obligations for vetted researchers and supervision fees.

The compliance work is not cosmetic. Systemic risk assessment under the Digital Services Act requires an organization to identify how its own design choices create risk, document mitigations and submit to external audit of both. For a generative system, that means answering questions the Act's drafters were not thinking about when they wrote it: what a ranking obligation means when there is no ranked list, how you audit a recommender system that is a language model, what electoral integrity mitigation looks like for a system that generates rather than surfaces.

The four month clock matters for anyone building on top of these services. Compliance obligations landing on a platform tend to propagate to the businesses that depend on it, through changed defaults, new content restrictions, altered API behavior and additional contractual terms. Enterprises with production deployments in Europe should assume that something about the behavior or the terms of the underlying service will change before January 2027, and should be asking their vendors now what specifically is expected to change rather than being told after the fact.

There is also a strategic read for anyone tracking where AI regulation is actually going. Europe already has the AI Act, which is the instrument built for this technology. What the Commission did here was reach for a different instrument, one that already has enforcement machinery, established audit practice and a track record of extracting behavioral change from very large platforms. Regulators do not need a new law to reach a new technology when an existing one fits closely enough. Companies planning their compliance posture around AI specific legislation alone are planning for half the surface area.

Reddit and Roblox drew the platform designation on the same day, which is a reminder that the Commission is expanding this net broadly rather than singling out AI.

AI Infrastructure Story 7 of 12

Anthropic Reported to Sign a $35 Billion Cloud Deal With Nvidia Backed Lambda

Bloomberg reported that Anthropic signed a $35 billion cloud computing agreement with Nvidia backed Lambda. Reuters reported that the agreement is tied to a data center in Nueces County, Texas developed by Hut 8. Anthropic has not published the arrangement in its own newsroom, so the figures here rest on press reporting rather than company disclosure, and executives should weigh them accordingly.

Taken at the reported scale, the deal belongs to a category of transaction that barely existed two years ago: a model developer committing tens of billions of dollars to a single compute supplier before the capacity it is buying has been built. Anthropic has now assembled compute commitments across several such relationships, spanning hyperscalers, chip designers and specialized providers, and the aggregate has moved well past what any one balance sheet comfortably carries.

Three structural features of this arrangement are worth understanding, because variations on them are showing up across the industry. The first is disintermediation of the hyperscalers. Lambda is a specialized AI cloud provider, not one of the three large public clouds, and a deal of this size routed through such a provider signals that model developers are willing to build supply relationships outside the platforms that historically owned enterprise compute. The second is the vendor financing loop. Nvidia backs Lambda, Lambda buys Nvidia hardware, and Anthropic's commitment underwrites the purchase. That circularity is not inherently improper, and it is now common enough across the sector that it has become one of the central questions in any honest assessment of how durable AI infrastructure demand really is. The third is the physical anchor. This is a specific data center in a specific Texas county, developed by a company whose origins are in bitcoin mining, which is where a meaningful share of the industry's convertible power and land sits.

For enterprise leaders, the relevance is less about Anthropic and more about the market you are buying into. Multi year, multi billion dollar prepurchase commitments by model developers are absorbing capacity years ahead of delivery. That is the same dynamic visible in Dell's $95 billion backlog, viewed from the other side. When you negotiate for inference capacity, dedicated capacity or committed throughput, you are negotiating for the residual after these commitments are met.

It also raises a question worth putting to any frontier vendor in a procurement conversation: what portion of your compute is contracted, from whom, on what schedule, and what happens to my service if a counterparty cannot deliver on time. Those are ordinary supply chain questions. They are asked routinely about semiconductors and rarely about model providers, even though the dependency is now comparable in size and considerably less transparent.

Absent confirmation from Anthropic, treat the specific numbers as reported rather than established.

Enterprise AI Story 8 of 12

OpenAI Connects ChatGPT to Epic Health Records With Read Only Access for Clinicians

OpenAI launched an Epic electronic health record integration for ChatGPT for Healthcare on September 1, 2026, giving clinicians the ability to bring authorized patient records into ChatGPT and, in some deployments, to reach ChatGPT from inside the record workflow itself. The integration provides read only access to authorized patient information. Nothing is written back to the chart.

That constraint is the design decision that makes the rest of it possible. Read only access removes the entire class of risk that has stalled clinical AI deployments for years, which is a system that alters the medical record based on a generated inference. What remains is synthesis: pulling appointment notes, laboratory results, medications and specialist documentation into one place so a clinician can review a history, identify what changed and prepare for a visit without assembling the picture manually across a dozen screens.

OpenAI paired the launch with a Healthcare Public Data plugin connecting nine official sources, among them ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed and PubMed, which supports trial eligibility screening and medication questions against authoritative references rather than model recall. The company reported that physicians rated 99.1% of responses safe across 27 clinical use cases.

Availability is tiered by compliance posture. ChatGPT for Healthcare customers can request administrator enablement. ChatGPT Enterprise customers with Regulated Workspaces go through sales. Individual clinicians in the United States get the public data plugin but not the record integration. Customers with an applicable Business Associate Agreement can use the features in workflows subject to United States health privacy law.

For executives in regulated industries generally, and not only healthcare, this launch is a template worth studying. The pattern is: constrain the write path to zero, integrate against the system of record rather than asking users to copy data out of it, publish a safety evaluation with a number attached, and gate availability by the customer's contractual and compliance status rather than by subscription tier alone. That combination is how a frontier vendor enters a regulated workflow without asking the customer to accept an unbounded risk.

It is also worth noting what the safety number does and does not say. A 99.1% safety rating across 27 clinical use cases is a meaningful disclosure, more than most vendors publish, and it is a rating by physicians of response safety rather than a measure of clinical outcomes. The remaining fraction of a percent, in a system used across millions of encounters, is not a rounding error, which is precisely why the read only constraint carries so much weight.

The competitive dimension is that Epic already has generative AI partnerships in place, so this arrives as an additional layer on a chart that has several. Health system chief information officers now face a governance question that will become familiar in other industries: when three vendors offer overlapping AI assistance inside one system of record, who owns the evaluation, the audit trail and the answer when two of them disagree.

Funding & Investment Story 9 of 12

Tencent Backed Enflame Prices Its Shanghai Listing as China's AI Chip Cohort Completes Its Run

Enflame Technology priced its Shanghai STAR Market initial public offering at 142.18 yuan per share. Caixin reported that the offering raised about 6.12 billion yuan and implies a valuation of roughly 61.2 billion yuan, and that the company's revenue rose from 301 million yuan in 2023 to 990 million yuan in 2025. Caixin also reported that Tencent accounted for 83.8% of Enflame's 2025 revenue.

That last figure is the one to carry into any analysis of China's domestic AI chip sector. Enflame is a chip designer whose growth is real and whose customer base is, by its own disclosure, overwhelmingly a single strategic backer. Tencent is both a major shareholder and the source of the great majority of revenue. Investors buying this listing are buying a bet that the company can convert a captive relationship into a broad merchant business, in a market where Huawei and several well funded peers are pursuing the same customers.

Enflame is the last of the group Chinese media call the four little dragons of AI chips to reach the public market, following Moore Threads, Biren Technology and MetaX. All four grew in the space created by United States export restrictions on advanced Nvidia hardware and by Beijing's push toward domestic self sufficiency. Their collective arrival on the STAR Market within a compressed window is the clearest available measure of how much domestic capital has been mobilized behind that policy.

Enflame's technical bet differs from its peers. Rather than pursuing general purpose graphics processors, it has built around domain specific architecture aimed at targeted workloads such as inference. That is a defensible strategy in a market where inference volume is growing faster than training and where matching Nvidia's general purpose performance is not realistic under current constraints. It is also a narrower strategy, and narrower strategies depend on the target workload staying where you expected it.

For Western executives, three observations are useful. First, the revenue trajectory, roughly tripling over two years, is real growth from a small base, and the valuation implied by the pricing sits well above what those revenues would support in a market pricing on fundamentals alone. This is strategic capital expressing a national priority. Second, a domestic Chinese inference chip ecosystem reaching public market scale changes the medium term calculus for any company with China operations, because domestically supplied inference capacity becomes a viable alternative to constrained imports. Third, the concentration risk visible in Enflame's revenue is a feature of the entire cohort, not a quirk of one company, because these businesses grew inside the procurement budgets of a handful of large platforms.

The demand signal is unambiguous. The South China Morning Post reported that the retail tranche of the offering was oversubscribed about 4,073 times. A number like that says less about Enflame's fundamentals than about how Chinese investors are positioning around domestic semiconductor self sufficiency as a durable policy commitment rather than a temporary response to export controls.

Industry Dynamics Story 10 of 12

Palo Alto Networks Buys Console to Push Agentic Automation Into Security Operations

Palo Alto Networks announced it has acquired Console, described as an AI native platform that enables agentic capabilities, and said the technology will deepen agentic capabilities in Cortex, supporting teams as they investigate signals, prioritize work and take action across their environment. Terms were not disclosed.

The framing in the announcement is worth reading closely. Nikesh Arora, chairman and chief executive officer of Palo Alto Networks, said security operations can no longer be about managing dashboards and queuing tickets just to help humans work faster. Palo Alto Networks identified Andrei Serban as co founder and chief executive officer of Console, and quoted him saying people should be able to express an operational goal and have intelligent software handle the complexity.

Both statements are aimed at the same structural problem. The security operations center as an industry built it is a queue management system. Detection tooling generates signals, correlation reduces them somewhat, and analysts work a backlog that is larger than the staff available to work it. Every generation of tooling has promised to help analysts move through that queue faster. The argument being made here is that faster queue processing is the wrong objective, and that the work itself should be executed by software operating against stated goals.

Whether that argument holds is the open question, and it is the same question every agentic enterprise deployment faces. Security operations is a demanding proving ground for it. The actions an agent might take, isolating a host, revoking credentials, blocking traffic, disabling an account, are consequential and sometimes irreversible in operational terms. An agent that acts confidently on a false positive causes an outage. An agent that requires human approval for every action reproduces the queue it was meant to eliminate. The design space between those failure modes is narrow and is where the real product work sits.

The transaction also reflects the consolidation logic now shaping enterprise software generally. A platform vendor with an installed base and a distribution channel buys a capability rather than building it, because the window in which agentic security automation is a differentiator is short and the engineering talent is concentrated in small companies. Expect more of these. The security tooling market has a large number of well funded startups building agentic layers on top of existing detection stacks, and a small number of platform vendors who can acquire them and put the capability in front of thousands of existing customers immediately.

For security leaders, the practical questions to bring to a vendor conversation are unchanged by the acquisition and are worth writing down now: which actions will an agent take without approval, what is the audit trail, how is a wrong action detected and reversed, and who is accountable when it is not. Any answer that leads with autonomy and treats those questions as implementation details is not ready for your environment.

AI Infrastructure Story 11 of 12

iPronics Raises $125 Million With Nvidia Participating to Rewire AI Clusters With Light

iPronics raised $125 million in a Series B round announced on September 2, 2026, co led by Maverick Silicon and Light Street Capital. NVIDIA joined the round, alongside participants including Triatomic Capital, Bosch Ventures, Catalight Capital, the European Innovation Council Fund and Amadeus Capital Partners. The company said total funding now stands at $177 million.

iPronics makes silicon photonics based optical circuit switching for AI data centers. Its product, a rack mounted optical switch called iPronics ONE, allows the physical connectivity between accelerators in a cluster to be reconfigured programmatically in real time. The company was founded in 2019 as a spin off from the Universitat Politecnica de Valencia, is headquartered in Valencia, Spain, and has opened an office in Santa Clara, California.

The problem being addressed is unglamorous and increasingly expensive. In a large training or inference cluster, the network topology connecting accelerators is fixed at build time, while the communication pattern of the workload running on it is not. Different model architectures, different parallelism strategies and different phases of the same job want different topologies. When the fixed network does not match the workload, accelerators sit idle waiting on data. Optical circuit switching lets the topology follow the workload rather than the other way around, and it does so in the optical domain, which avoids the power cost of converting signals to electrical form and back at every hop.

The financial case is straightforward once framed correctly. Utilization is the denominator under every AI infrastructure cost calculation an enterprise runs. An underused cluster of expensive accelerators costs the same to own as a fully loaded one, and the whole of that difference lands in the cost per token, per training run or per inference. Networking has quietly become one of the largest determinants of that number, and one of the least discussed in executive conversations about AI capital expenditure.

Nvidia's participation is the detail that should register. Nvidia has an obvious interest in anything that raises the effective throughput of clusters built around its accelerators, and its investment pattern is a reasonable leading indicator of where the company sees the bottleneck moving next. That the bottleneck is being addressed in optics, and by a European startup with roots in a Valencia university lab, is also a data point about where advanced photonics expertise actually sits.

For enterprises building or leasing dedicated AI capacity, the near term implication is a question to ask providers: what is the measured accelerator utilization on the capacity I am buying, and what is the network topology doing when it drops. Most capacity contracts are priced on nominal accelerator count and hours, which is a measure of what you are paying for rather than what you are getting. Technologies like this one are aimed squarely at the gap between the two, and the vendors deploying them first will have a real cost advantage they can either pass through or keep.

AI Business Models Story 12 of 12

Perplexity Puts a Privacy Gate on the Mac and Keeps Sensitive Work Off the Cloud

Perplexity launched hybrid compute on Mac on September 1, 2026, splitting AI work between cloud models and models running locally on the user's machine. It works on any Apple silicon Mac running macOS 15 or later with at least 24GB of unified memory, and it is available to Perplexity Pro, Max and Enterprise subscribers.

The mechanism is a privacy gate that runs on the device. An on device classifier identifies sensitive material such as names, addresses, account numbers and secrets, and the gate can then mask those details, keep the work local, refuse the action outright, or ask the user for consent before anything leaves the machine. Frontier reasoning, web search and planning continue to run in the cloud. Private files, sensitive information, on device actions and client data stay local. Three local models ship with it, Gemma 4 E4B, Qwen3.6 35B-A3B, and a Perplexity model.

The enterprise controls are what make this more than a consumer privacy feature. Administrators on Enterprise plans can set organization wide rules for what must stay on the Mac, what may be masked before cloud use and what requires explicit user approval, and can review logs of when and what information left a device.

This is a direct answer to the objection that has blocked more AI deployments in professional services, law, healthcare and finance than any other: that using a capable assistant on real client work requires sending that work to a third party. The usual responses have been contractual, in the form of data processing terms and retention commitments, or architectural in a heavier sense, such as private deployments. A device level gate with administrator policy and an audit log is a third option, and it is a considerably cheaper one.

The tradeoff is capability. A model running inside 24GB of unified memory is not a frontier model, and the tasks routed locally will be handled less well than the tasks routed to the cloud. What the design gets right is that the split follows sensitivity rather than difficulty, which is the correct axis for a governance control even though it is the wrong axis for output quality. Users will notice when the sensitive path is the weaker path.

For technology and risk leaders, the useful question this raises is not whether to adopt this particular product. It is whether your organization's AI data policy is currently written as a blanket rule because a blanket rule was the only thing enforceable. If sensitivity aware routing with administrator policy and logging is now available at the endpoint, a policy that says no client data in AI tools may be leaving significant productivity on the table for a control that was chosen out of necessity rather than judgment. That is a policy worth revisiting deliberately, with the security team in the room, rather than letting individual users resolve it on their own.