Altman Pumps AI Brakes.

Mazen Al Sabbagh

— 6 min read
Share

The week's most salient development isn't a new model, but a strategic re-evaluation from its architects: the call for deliberate deceleration. This pivot by key industry players, juxtaposed against relentless market expansion and escalating geopolitical tensions, forces a recalibration of capital allocation and foundational business architecture.

TL;DR

  • Altman signals caution: OpenAI's CEO informed staff that the lab is exploring a slowdown in cutting-edge AI development, ideally in concert with other leading labs, marking a significant shift for a previously staunch accelerationist.
  • DeepSeek's new release sparks debate: DeepSeek-V4.1-Flash launched as a 552B-parameter, efficiency-focused open model, arriving mere days after Washington accused Chinese labs of industrial-scale model theft.
  • Memory costs surge: Samsung has increased DRAM prices by up to 60% since September, as the global AI build-out creates a severe supply bottleneck, leading analysts to project a memory "supercycle" for 2026.
  • Inference silicon attracts capital: Positron, an inference-chip startup, secured $875M at a $5B valuation—a fivefold increase in seven months, despite its "Asimov" chip not having taped out yet.
  • Grok 4.7 remains unreleased: Elon Musk's anticipated launch window for Grok 4.7 passed without a public model page, API, or benchmark, leaving its status confined to social media mentions.

Lead Story: Altman Pumps AI Brakes

Sam Altman spent two years publicly articulating a vision where the path to AI safety lay in accelerated development. This week, his public stance underwent a notable shift, hardening from a tweet into concrete internal discussion.

The initial indicator emerged from models themselves. Following OpenAI's disputed Navier-Stokes proof, Altman remarked that the world "has extremely capable models now," expressing surprise at the magnitude and speed of such a result. He framed this as "the strongest evidence yet of the urgency" to pace progress, prioritizing safety.

The shift became actionable. Altman informed employees in a company-wide meeting that OpenAI is considering a slowdown in the development of its most advanced systems. This contemplation includes a hope for industry-wide participation, though he acknowledged that not all labs may comply. This isn't an idle suggestion: OpenAI had already paused internal Astra work on August 7 due to security concerns, delaying the model's release on September 1 to enhance its resilience against cyber misuse.

While no public releases have been canceled and no specific timeline announced, industry observers are interpreting the move. CAIS director Dan Hendrycks called it "a responsible course correction." Conversely, a former OpenAI policy staffer viewed it as "positioning ahead of regulatory deadlines rather than a genuine strategy shift," underscoring the strategic ambiguity.

The timing is not coincidental. It arrives in the same week an Anthropic researcher, Jacob Coxon, publicly resigned, warning that labs are "gambling with our lives." Anthropic's own alignment lead, Evan Hubinger, echoed this, citing over 10% odds of AI-induced human extinction within a decade.

This discourse is now actively shaping Washington's approach. Lawmakers are mobilizing: Senators Thune and Klobuchar are drafting legislation for mandatory pre-deployment government testing, Senator Cruz is referencing catastrophic-risk bills, and Representative Anna Paulina Luna has requested a special session. Bernie Sanders is convening a bipartisan Senate briefing on September 16 featuring Geoffrey Hinton and Max Tegmark, pointedly inviting risk researchers over lab executives. Meanwhile, former President Trump maintains a dismissive stance on extinction scenarios, advocating for minimal regulatory intervention.

The through-line remains clear: capability is outpacing verification. Paradoxically, the very architects of this accelerating capability are now advocating for deceleration. The central question is whether this internal "considering" will translate into a slower cadence, or if it represents the most judicious posture to adopt in anticipation of looming regulatory deadlines.

In Other News

DeepSeek ships V4.1-Flash, adding geopolitical friction. DeepSeek introduced V4.1-Flash on September 10, a 552B-parameter mixture-of-experts model leveraging a new "causal encoder-decoder" design for significant efficiency gains, requiring only ~8B active parameters and supporting native multimodal input. Per its documentation, it outperforms the flagship V4-Pro while reducing KV-cache memory requirements by approximately 75%. Its release, days after a US intelligence advisory accused DeepSeek and five other Chinese labs of distilling American frontier models, elevates its market entry to a clear geopolitical signal, impacting global AI architecture and intellectual property discussions.

The AI boom is repricing memory, signaling a supercycle. Samsung has increased memory-chip prices by up to 60% since the beginning of September, with a 32GB DDR5 module now costing $239, up from $149. The company has issued warnings of industry-wide shortages extending into 2026. SK Hynix is reportedly at near-complete sell-out for HBM and DRAM for the current year. This supply crunch demonstrates how the datacenter arms race is directly translating into higher costs across the entire bill of materials.

Positron raises $875M, intensifying the inference stack challenge. Reno-based Positron AI closed an $875M round on Thursday, valuing the company at $5B—a fivefold increase from its $1B valuation post-February's Series B. Its "Asimov" chip is designed to utilize commodity LPDDR5X memory, enabling up to 2,304GB per chip for inference, despite not having taped out yet. This round stands as the clearest signal among a flurry of chip investments this week that capital is increasingly flowing into inference-focused silicon, mirroring last week's Qualcomm-Amazon custom-chip pact and reinforcing a shift in the compute architecture.

Beijing formally rejects US distillation accusations. China's Commerce Ministry labeled the US "industrial-scale distillation" accusations as "groundless" and "without legal basis," arguing that distillation is a neutral, widely used technique. The ministry reiterated warnings of "resolute countermeasures" if the claims are leveraged to suppress Chinese firms, though no concrete retaliation had materialized by Thursday. The US advisory itself reportedly discussed "AI poisoning"—tampering with responses to suspected scrapers—all ahead of a Trump-Xi meeting scheduled for around September 24, intensifying global market and IP tensions.

X / Social Pulse

  • Evan Hubinger (Anthropic): The alignment lead openly endorsed Jacob Coxon's warnings, stating, "Jacob is correct here, we really do earnestly believe AI could kill all humans," and placing extinction odds above 10% within the decade. The internal debate is now fully external.
  • Terence Tao: While acknowledging the human Navier-Stokes work as "remarkable," the mathematician warned of a "decoupling between getting answers and getting understanding," raising concerns about a "meaningless production quota" driven by synthetic results.
  • Buckmaster vs. Bubeck: NYU's Tristan Buckmaster alleges that OpenAI's Sébastien Bubeck pressured him to remove a co-author, an accusation Bubeck denies, claiming no sight of private work. The ongoing credit dispute highlights the intense reputational and IP incentives in frontier research.
  • Elon Musk: Continued to promote Grok 4.7, citing 2.1T parameters trained on SpaceX data, yet the promised model remains unreleased, maintaining a speculative valuation for its potential.
  • David Sacks: The PCAST chair publicly reiterated his position that Anthropic's IPO "must be paused" pending an investigation into Coxon's claims, injecting a new political and ethical vector into the company's market debut.

One to Watch

Grok 4.7. Musk's self-imposed September 11-12 launch window elapsed without a release; xAI's model list still tops out at Grok 4.6. Consider it unshipped until an official page or API emerges. Gemini 3.5 Pro remains absent after multiple delays. Anthropic's IPO is now pushed to mid-October, with a $15B revolver secured but no public S-1 filing on EDGAR—and Sacks' pause call adds a new political dynamic. Also, monitor whether US-China AI talks materialize before the September 24 summit, and the outcome of independent refereeing for the Navier-Stokes proof.

Quick Hits

  • Anthropic secured a $15B revolving credit facility but delayed its IPO to mid-October; the public S-1 remains outstanding.
  • Nvidia's reported $17-20B Groq license deal is under DOJ scrutiny, examining potential circumvention of merger review processes.
  • Bartz v. Anthropic author payouts are expected around September 17, with approximately 120,000 rightsholders having filed claims.
  • The EU is treating OpenAI's DseWiki incident report as a critical first test for the AI Act's serious-incident disclosure regulations.
  • Positron was not alone: Kepler Computing raised $468M and UniPat AI secured ~$300M, signifying a particularly active week for chip and infrastructure funding rounds.

The week's center of gravity shifted definitively from celebrating model capabilities to an urgent interrogation of who, if anyone, will manage the pace of development. Tellingly, the loudest new voice for caution is the individual who built the accelerator. Capability continues to outrun verification. The critical question now is whether Washington or the labs themselves will be the first to truly engage the brake, redefining the incentives and architecture of this market.

Sources

Lock in. M. mazen@thorterminal.com

Read more