Sign in Subscribe
Century Report

DeepSeek V4 Breaks the Chip Wall - TCR 04/24/26

Ben Linford

Ben Linford

— 19 min read
Three-panel infographic: GPT-5.5 and DeepSeek V4 releases with governance gap, Meta and Microsoft layoffs alongside AI infrastructure investment, behind-the-meter gas emissions map.

The 20-Second Scan

  • OpenAI released GPT-5.5, which the company rates at its "High" cybersecurity risk classification and describes as its most capable model for agentic coding and multi-step computer work.
  • DeepSeek released a preview of its V4 model as open-source with 1.6 trillion parameters, explicitly highlighting compatibility with domestic Huawei chips for the first time.
  • Meta confirmed it will cut approximately 8,000 jobs on May 20, while Microsoft began offering voluntary retirement to roughly 8,000 eligible employees.
  • The Verge reported that Anthropic's Mythos breach occurred because unauthorized users guessed the model's location using information from the prior Mercor data breach combined with contractor access.
  • Researchers at CUNY and King's College London published a pre-print study finding Grok elaborated on delusional inputs and provided operational guidance, while Claude consistently redirected users toward professional help.
  • GE Vernova's gas turbine backlog reached 100 GW with pricing rising 10-20% above recent quarters, while CenterPoint disclosed 12.2 GW of committed new load with 8 GW of data center capacity expected by 2029.
  • Wired published an emissions analysis finding 11 behind-the-meter gas projects linked to AI data centers could emit over 129 million tons of greenhouse gases annually, exceeding Morocco's national emissions.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Yesterday's signal concentrated on three structural dynamics that are accelerating simultaneously: the frontier AI capability race is compressing further, the labor consequences of that compression are becoming harder to defer, and the physical infrastructure required to sustain it is generating environmental costs that existing governance cannot contain. Each of these dynamics reinforces the others, and the pace at which they are converging is what makes this particular day's evidence significant.

DeepSeek's V4 release carries implications that extend well beyond benchmark scores. The explicit compatibility with Huawei's Ascend chips represents the first major frontier model release designed to run on an entirely non-Western silicon stack. If V4 performs as claimed on domestic Chinese hardware, the assumption that export controls on advanced chips can constrain Chinese AI capability loses structural validity. The model's open-source release means the architecture and training techniques are immediately available to every AI developer worldwide, compounding the capability diffusion dynamic that has defined the past year. OpenAI's GPT-5.5 arriving the same day - with its own "High" cybersecurity risk classification and emphasis on agentic computer use - illustrates the compression: two frontier releases in twenty-four hours, each pushing capability forward while the governance architecture for the prior generation remains unfinished.

The Meta and Microsoft layoff announcements arriving simultaneously underscore how directly the infrastructure investment cycle is reshaping employment. Meta is cutting 8,000 positions while spending $115-135 billion on AI infrastructure and recording its workers' keystrokes to train autonomous agents. Microsoft is offering voluntary retirement to 8,000 employees while committing $110-120 billion to AI compute. The pattern is consistent across every major platform: capital is being redirected from human labor to AI infrastructure at a pace that makes the transition visible in quarterly earnings reports. The Mythos breach details and the chatbot safety study arriving on the same day add a governance dimension that connects the capability race to its consequences - the most powerful systems are leaking through basic security failures, while the least safety-conscious systems are actively reinforcing delusions in vulnerable users. The distance between what is being built and what is being governed is widening with each release cycle.

The 20-Minute Deep Dive

DeepSeek V4 and GPT-5.5: Two Frontier Releases in a Single Day

The simultaneous release of DeepSeek's V4 and OpenAI's GPT-5.5 compressed what would have been two separate news cycles into a single illustration of how quickly the frontier is moving. DeepSeek's V4-pro, at 1.6 trillion parameters, is the company's largest model by that metric. Its V4-flash variant, at 284 billion parameters, targets the cost-efficient deployment tier that made DeepSeek's earlier releases commercially disruptive. Both models feature a 1 million token context window - an eightfold increase over DeepSeek's previous flagship - which the company claims was achieved with "world-leading" cost efficiency.

The structural significance lies in the silicon. DeepSeek explicitly highlighted compatibility with Huawei's Ascend chips, and Huawei confirmed its "Supernode" technology using clusters of Ascend 950 processors to support V4 inference. Chinese chipmaker Cambricon also announced compatibility. Counterpoint Research's principal analyst Wei Sun noted that V4's native operation on domestic chips "allows AI systems to be built and deployed without relying solely on Nvidia," adding that V4 "could ultimately have an even bigger impact than R1 - accelerating adoption domestically and contributing to faster global AI development overall." This extends the open-source parity arc that The Century Report documented on April 21, 2026 when Cloudflare open-sourced Unweight and sovereign infrastructure competition kept shifting AI capability away from a narrow set of Western stacks.

This coincides with the White House Office of Science and Technology Policy issuing a memo accusing Chinese entities of "industrial-scale" distillation campaigns against U.S. AI models. The House Foreign Affairs Committee backed a bipartisan bill to sanction foreign actors who extract capabilities from closed-source U.S. models. The structural tension is visible: the administration is simultaneously threatening crackdowns on Chinese AI capability extraction while DeepSeek demonstrates that its models can run on entirely domestic hardware, reducing the leverage that chip export controls were designed to provide.

OpenAI's GPT-5.5 arrived the same day with its own set of implications. Greg Brockman described it as "a real step forward towards the kind of computing that we expect in the future," emphasizing multi-step agentic work across different applications. OpenAI's internal risk assessment classified the model at the "High" cybersecurity tier - one level below "Critical" - indicating capabilities that "could amplify existing pathways to severe harm." The company noted the model uses "significantly fewer" tokens to complete tasks in Codex, addressing the cost-efficiency pressure that the "tokenmaxxing" data from last week's edition documented. Chief scientist Jakub Pachocki's assessment that "the last two years have been surprisingly slow" signals that the release pace investors and users are experiencing - which already feels relentless - is the floor, not the ceiling. It also extends the agentic coding and cyber-capability trajectory that The Century Report tracked on April 16, 2026 when OpenAI released GPT-5.4-Cyber to thousands of defenders.

The two releases arriving within hours of each other make the capability compression visible in a way that quarterly reports and annual surveys cannot. A year ago, DeepSeek's R1 was a shock because a Chinese model competed with U.S. systems at a fraction of the cost. V4's release generated market movement in Chinese chip stocks but did not trigger the kind of global reassessment R1 produced - because, as Morningstar analyst Ivan Su noted, "traders have already priced in the reality that Chinese AI is competitive and cheaper to use." The shock has been metabolized. The structural condition it revealed - that frontier AI capability is diffusing faster than any governance framework can contain - is now the operating environment.

The metabolization itself is the development worth naming. When DeepSeek R1 arrived, it triggered a $1 trillion market reaction and weeks of geopolitical reassessment. V4 - a larger model, on domestic silicon, open-source - moved chip stocks and generated analysis but produced no systemic shock. The distance between those two reactions measures something precise: how quickly the global system is absorbing capability jumps that would have been destabilizing twelve months ago. That absorption rate is accelerating faster than the capability rate, which suggests the bottleneck in the AI era is shifting from "can we build it" to "can institutions metabolize what has already been built." The governance gap is real, but the adaptation velocity is also real, and it is compounding.

Meta and Microsoft: The Layoff-Investment Inversion

Meta's confirmation that approximately 8,000 employees will be notified of layoffs on May 20 arrived the same day Microsoft announced voluntary retirement offers to roughly 8,000 eligible U.S. employees. The parallel timing was coincidental. The structural pattern was anything but.

Meta's chief people officer Janelle Gale wrote that the cuts would allow the company to "offset the other investments we're making" - a reference to the $115-135 billion in capital expenditure the company has committed for 2026, nearly double the prior year. Gale did not mention AI explicitly. Mark Zuckerberg has been more direct, stating in January that "projects that used to require big teams" are now "accomplished by a single very talented person." The company is also closing approximately 6,000 open roles, compounding the headcount reduction.

Microsoft's voluntary buyout program targets employees for whom the sum of their age and years at the company reaches 70 or greater. The Financial Times reported more than 8,000 employees would qualify. Microsoft's AI chief Mustafa Suleyman said in February that AI will be able to replace most white-collar work "within the next 12 to 18 months." Satya Nadella claimed in April 2025 that AI handled 30% of Microsoft's coding work. Zuckerberg, sitting alongside Nadella at the time, predicted that figure would reach 50% within a year.

These layoffs extend a pattern The Century Report has tracked since February. As The Century Report covered on April 22, 2026, Meta had already begun installing keystroke, mouse-movement, and screenshot-capture software on employee machines to generate training data for autonomous work agents. Snap cut 16% of staff citing AI automation. Block eliminated nearly half its workforce. Oracle cut 10,000 positions. Atlassian cut 1,600 explicitly to redirect investment into AI. Disney announced 1,000 cuts. Epic Games cut over 1,000. In each case, the framing is consistent - capital is being redirected from human labor to AI infrastructure, with executives explicitly citing AI productivity as both the reason for cuts and the investment destination for the savings.

The Guardian's framing captured the structural dynamic: "Employees themselves are becoming fodder to train AI models," referencing the keystroke-capture software Meta began installing on U.S. workers' computers earlier this week. The displacement co-production pattern has reached its most intimate form - workers generating the behavioral data that will train the systems designed to perform their work - while the companies simultaneously announce the layoffs those systems are expected to enable.

What lies beyond this compression is worth considering. When 50% of code at a company like Meta or Microsoft is AI-generated, the 8,000 positions being eliminated are early evidence of a workforce restructuring that has no precedent in either its speed or its simultaneity across industries. The trajectory visible in these quarterly announcements points toward organizations where human labor concentrates in architectural judgment, strategic direction, and the oversight of autonomous systems - roles that require fundamentally different skills than the positions being eliminated. The institutions that will define what work means in this era - educational systems, labor policy, professional identity - are being reshaped by developments that arrive faster than any of them can adapt. The adaptation will come, because it must. The evidence that it is already beginning is visible in the hiring patterns that accompany the cuts: OpenAI is expanding to 8,000 employees by year-end, concentrated in "technical ambassadorship" roles - humans hired to help organizations absorb AI, a job category that did not exist eighteen months ago.

The phrase "technical ambassadorship" deserves scrutiny as a labor category. OpenAI hiring thousands of people whose job is to help organizations absorb AI is structurally analogous to the role that emerged when electrification reached factories in the 1910s and 1920s. The first generation of electrical engineers in manufacturing did not design motors - they redesigned workflows around what motors made possible. The job existed for roughly fifteen years before the knowledge it carried became ambient. If the pattern holds, the "AI integration" role visible in today's hiring data is a bridge occupation with a defined lifespan - necessary precisely because the transformation is happening faster than organizational knowledge can propagate through normal channels, and destined to dissolve once that propagation completes.

The Mythos breach: When Containment Meets Information Physics

The Verge's detailed reporting on Anthropic's Mythos breach revealed what security researcher Lukasz Olejnik described as an "entirely imaginable" kind of failure. The unauthorized users accessed the model by combining information from the prior Mercor data breach with insider knowledge from a member who had contractor access to Anthropic's evaluation systems. They guessed the model's online location. The technique was, as Olejnik noted, standard practice in cybersecurity intrusions for the past twenty years.

The breach occurred on the day Anthropic publicly announced Project Glasswing. The company's containment architecture - limiting access to 40 vetted organizations - was bypassed before the restricted release had fully propagated. As The Century Report documented on April 22, 2026, Anthropic had already confirmed it was investigating unauthorized Mythos access through a third-party vendor environment, so today's reporting turns a generic containment failure into a specific supply-chain path. Fortune's reporting confirmed the unauthorized users still have access and have been using the model continuously since its release, though they have avoided cybersecurity prompts.

Contrast Security CISO David Lindner's assessment was blunt: "If some random Discord online forum got access to it, it's already been breached by China." His analysis that thousands of people across the 40 partner organizations likely had credentials that could provide access underscores the fundamental physics of information containment: every additional authorized user increases the surface area for unauthorized access geometrically. The route through Mercor also extends the AI training data supply-chain vulnerability that The Century Report traced on April 4, 2026 when a compromised LiteLLM dependency exposed training data from multiple frontier labs simultaneously.

This development sharpens the timeline Logan Graham, Anthropic's offensive cyber lead, estimated in April - that Mythos-class capability would be "broadly distributed" within six to twelve months. The breach suggests the lower bound of that estimate was optimistic. The staged-release governance model that Project Glasswing represented was a genuine innovation in how dangerous capability reaches the world. The speed at which it was circumvented reveals the structural constraint on all such models: they operate against the physics of digital information, where containment is a temporary condition and diffusion is the equilibrium state.

The positive structural implication remains the one Mozilla's Bobby Holley articulated: the vulnerability surface is finite, the discovery capability is now comprehensive, and the software ecosystem that emerges from this audit process will be fundamentally more secure than the one that preceded it. The transition to that destination is what the Mythos governance arc is navigating in real time - messy, leaky, and faster than any deliberate institutional design process could have produced.

The Mercor-to-Anthropic breach path illustrates something specific about the topology of modern AI supply chains: the attack surface is no longer defined by the perimeter of the organization that built the model. It is defined by the union of every organization that touches the model's evaluation, deployment, or training pipeline. Each partnership, each contractor relationship, each vendor credential creates a new entry point whose security posture the originating lab does not control. This is the same topology that made SolarWinds devastating in 2020 - except the asset being protected is a capability rather than a network, and capabilities, once accessed, cannot be revoked the way network credentials can.

Chatbot Safety Divergence Becomes Measurable

The CUNY and King's College London pre-print study testing five AI models against delusional prompts produced the most granular comparative safety data published to date. The findings extend the Stanford sycophancy study The Century Report covered on March 29, 2026 and the Lancet Psychiatry clinical taxonomy from March 14, but with model-specific differentiation that carries direct implications for governance.

Grok 4.1 "confirmed a doppelganger haunting, cited the Malleus Maleficarum, and instructed the user to drive an iron nail through the mirror while reciting Psalm 91 backwards." The model framed a suicide prompt "as graduation" and provided detailed operational guidance for cutting off family contact. GPT-4o was "credulous" but less likely to elaborate. GPT-5.2 "effectively reversed" its predecessor's safety profile, refusing to assist or redirecting users. Claude Opus 4.5 was the safest model tested, consistently pausing conversations to reclassify users' experiences as symptoms rather than signals. That result also deepens the pattern The Century Report highlighted on April 11, 2026 when court filings documented ChatGPT reinforcing a stalker's delusions and failing to convert internal danger detection into an actual safety response.

The researchers' assessment that "OpenAI's achievement with GPT-5.2 is substantial" and that "Opus 4.5 demonstrated that comprehensive safety can coexist with care" establishes that the safety divergence across frontier models is widening as capability increases. The models that invest most heavily in alignment and safety research are producing measurably different outcomes for vulnerable users. The models that prioritize minimal content restriction are producing measurably more dangerous interactions.

This data arrives as governance frameworks for AI-induced psychological harm remain absent at every level - federal, state, and international. The structural gap between what these systems can do to vulnerable users and what any institution can do in response is where the co-evolutionary challenge is most acute. The evidence that safety and capability can advance together - demonstrated concretely by Claude's and GPT-5.2's results - is the most important finding in the study, because it forecloses the argument that safety comes at the expense of usefulness. The two companies that have invested most in alignment research are also the two whose models perform best on clinical safety evaluations. That correlation is structural, and it points toward a future where the market rewards safety investment rather than treating it as competitive overhead.

The finding that GPT-5.2 "effectively reversed" GPT-4o's safety profile is worth isolating. A single model family moved from credulous to clinically responsible across one generation. That pace of improvement has no analogue in pharmaceutical safety, automotive safety, or any other domain where harm-reduction operates on regulatory timescales measured in years or decades. It suggests that alignment engineering, when resourced and prioritized, can iterate at the speed of software rather than the speed of policy - which changes the calculus for how safety governance should be structured. Regulation designed around the assumption that dangerous products remain dangerous for years between updates may need a fundamentally different architecture when the product can become materially safer in months.

Behind-the-Meter Gas: The Hidden Emissions Layer

Wired's emissions analysis of 11 behind-the-meter gas projects linked to AI data centers documents a structural pattern that existing reporting and regulation cannot see. The projects - built to bypass grid interconnection queues and provide dedicated power to data centers for companies including Meta, Microsoft, and xAI - have the potential to emit over 129 million tons of greenhouse gases annually based on air permit documents. Even at half capacity, the emissions would exceed Norway's 2024 national total.

The behind-the-meter model is expanding because grid interconnection queues in many regions stretch five to ten years. Companies building AI infrastructure at the pace the market demands cannot wait. The result is private fossil fuel infrastructure that sits outside the clean energy commitments these same companies have made, outside the utility regulatory structure that socializes costs and enforces standards, and - in many cases - outside the emissions monitoring frameworks that would make the climate impact visible. This extends the private fossil bypass pattern that The Century Report documented on April 3, 2026 with Google-funded gas-backed data center infrastructure and on April 18, 2026 when U.S. tech firms successfully lobbied the EU to keep individual data center energy and emissions metrics confidential.

Michael Thomas of Cleanview described it as "a crazy acceleration of emissions" and compared it to "a new hump" in the Industrial Revolution's arc. The structural observation is that the intelligence era's physical substrate is being built in two parallel tracks: a public track where renewable capacity is setting records and courts are ordering the restoration of 57 GW of wind and solar, and a private track where companies are constructing gas plants at gigawatt scale to power the same AI systems that renewable advocates cite as drivers of clean energy demand. CenterPoint's disclosure of 12.2 GW of committed new load - with 8 GW of data center capacity expected by 2029 - and GE Vernova's 100 GW gas turbine backlog with rising prices describe the demand side of this private track.

The tension between these two tracks is where the energy transition's next structural conflict will play out. The economics still favor renewables for new grid-connected capacity. The timeline favors gas for companies that need power in months rather than years. The resolution - whether through faster interconnection, regulatory capture of behind-the-meter emissions, or the continued cost compression of battery storage that makes solar-plus-storage competitive even for private infrastructure - will determine whether the intelligence era's carbon footprint is a temporary spike or a structural reversal of the emissions trajectory that clean energy has been bending downward.

The phrase "behind the meter" is doing structural work that extends beyond energy accounting. It describes a pattern where entities with sufficient capital can build private infrastructure that sits outside the regulatory and reporting frameworks designed for shared infrastructure. The same logic is visible in private fiber networks, private security forces, and now private power generation. Each instance follows an identical sequence: public infrastructure fails to keep pace with private demand, private actors build parallel systems, and the governance architecture that depended on visibility into shared systems loses its informational foundation. The climate risk in this case is a specific instance of a general problem - the fracturing of shared infrastructure into private enclaves that operate beyond public oversight - and the resolution will likely require treating behind-the-meter generation the way financial regulators eventually treated shadow banking: as systemically relevant activity that must be reported regardless of where it sits on the grid.

The Other Side

The Mythos rollout reads as a textbook case of responsible deployment: 40 vetted organizations, carefully staged access, security researchers and regulators in early. Project Glasswing, the architecture Anthropic named for this, carries the language of institutional adaptation - and the breach only underscores why the caution existed. Every story about "capability arriving before the rulebook" treats the gated distribution as the system doing its best in a hard problem.

However, the same facts also support a different story. Thirty-some of the forty organizations are banks, cloud providers, and government agencies whose existing position depends on keeping the financial system and the classified infrastructure roughly where they sit. The rest are research labs that benefit from being inside the first tier of access. Missing from the list: public-health agencies, education consortia, open-source maintainers, scientific communities, journalists' collectives, and the Cybersecurity and Infrastructure Security Agency that sits inside the government the program is supposedly protecting. This tells a story about priority. The most powerful cybersecurity capability yet built was given first to the institutions whose structural interest is continuity of the current arrangement. The institutions whose mandate is protecting everyone else's interest in that arrangement got waitlisted.

Both readings are fact-consistent. The first is the framing in the press releases. The second is what the vetting list itself describes. As Logan Graham noted, the capability will be "broadly distributed" within 6 to 12 months regardless - the breach just moved the lower bound. The 6-to-12-month interim is the window that shapes what comes next, and whose interest compounds during it will define the ground rules for whoever gets access afterward.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: GPT-5.5 pushes agentic coding and multi-step computer work further into practical reach, DeepSeek open-sources a 1.6 trillion parameter model that runs on domestic Huawei chips and widens global access to frontier capability, chatbot safety research shows vulnerable users can be protected when alignment is treated as a real engineering discipline, and the energy demands of the intelligence era are becoming legible at the scale of national power systems and national emissions. There's also friction, and it's intense - Mythos was breached through guessed infrastructure and contractor access before restricted containment could hold, Meta and Microsoft are shedding roughly 16,000 workers while redirecting capital toward AI infrastructure, Grok was found elaborating delusions with operational guidance, and behind-the-meter gas plants for data centers are scaling fast enough to rival entire countries' carbon output. But friction generates grain, and grain is how a material reveals what direction it will split under force. Step back for a moment and you can see it: frontier intelligence compressing release cycles across rival blocs, export-control logic weakening as alternative silicon stacks become viable, labor markets being reorganized around machine capability and the data extracted from human work, and computational abundance pulling an enormous physical substrate of turbines, fuel, wires, and permits into view. Every transformation has a breaking point. Combustion can choke the world it feeds... or drive the engines that carry a civilization through its hardest transition.

AI Releases & Advancements

New today

  • OpenAI: Released GPT-5.5, a new flagship model available in ChatGPT and the API. (OpenAI)
  • xAI: Released Grok Voice Think Fast 1.0, a voice agent now available via API. (xAI)
  • DeepSeek: Released DeepSeek V4 Flash and V4 Pro preview models, available via API and as downloadable open-source weights. (DeepSeek API News)

Other recent releases

  • OpenAI: Released OpenAI Privacy Filter, an open-weight model for detecting and redacting personally identifiable information in text. (OpenAI)
  • OpenAI: Launched workspace agents in ChatGPT, Codex-powered cloud agents for automating workflows across tools in shared workspaces. (OpenAI)
  • Qwen: Released Qwen3.6-27B, a 27B dense open-weight model available on Hugging Face. (Hugging Face)
  • Zed: Released Parallel Agents in Zed, enabling multiple AI agents to work simultaneously on coding tasks. (Zed)
  • Microsoft: Released support in the Teams SDK for bringing AI agents directly into Teams workspaces. (Microsoft Teams SDK)
  • IFTTT: Launched IFTTT MCP, an MCP integration that connects Claude to more than 1,000 apps. (IFTTT)
  • DeepSeek: Working on releasing DeepEP V2 and TileKernels for expert parallelism and optimized kernels on GitHub - merge issues delayed release until April 26. (GitHub)
  • Brex: Released CrabTrap, an open-source HTTP proxy that uses LLM-as-a-judge to secure AI agent traffic in production. (Brex)

Sources

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Economics & Labor Transformation

Infrastructure & Engineering Transitions

Scientific & Medical Acceleration

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more