Permanent daily edition
The Open-Weight Efficiency Shock: DeepSeek-V3.5 Drops Frontier Token Economics to $0.14/M as US Finalizes Binding 10^26 FLOPs Reporting and TSMC Advanced Packaging Sells Out Through 2027
The Tuesday, September 15, 2026 New York morning edition, preserved with its cutoff, direct evidence, reader-first briefing, professional detail and audit appendix.
Executive summary
The Open-Weight Efficiency Shock: DeepSeek-V3.5 Drops Frontier Token Economics to $0.14/M as US Finalizes Binding 10^26 FLOPs Reporting and TSMC Advanced Packaging Sells Out Through 2027
DeepSeek-V3.5 cuts frontier API pricing by 80% with Multi-Head Latent Attention; the US Department of Commerce establishes binding dual-use reporting thresholds under the Defense Production Act; and TSMC packaging commitments lock in 2027 compute allocations.
Plain-English picture: The artificial intelligence landscape on September 15, 2026, is shaped by a stark divergence between algorithmic efficiency and physical hardware limits. DeepSeek released DeepSeek-V3.5, a 671-billion-parameter Mixture-of-Experts model that routes to just 37 billion active parameters per token. By combining fine-grained expert routing with Multi-Head Latent Attention (MLA), the model compresses the Key-Value (KV) cache by 93%, enabling frontier performance—82.4% on AIME 2026 and 49.6% on SWE-bench Verified—at an API price point of $0.14 per million input tokens. In Washington, the Department of Commerce issued binding reporting mandates under the Defense Production Act for training runs exceeding 10^26 FLOPs, imposing civil penalties of up to $250,000 per day for undisclosed frontier clusters. In hardware manufacturing, TSMC confirmed 68.2% yields on 2nm GAAFET silicon, but revealed that advanced packaging capacity (CoWoS-L) is 100% committed through 2027, with Nvidia and Apple taking 78% of total volume. Meanwhile, cybersecurity researchers documented Speculative Leak, a microarchitectural cache side-channel allowing co-tenants on shared GPUs to reconstruct private prompts during speculative decoding.
Decision-ready intelligence
4 developments that matter most
Facts, interpretation and recommended actions are separated. Quiet days are not padded to a fixed number of items.
Frontier Open-Weight Models & Algorithmic Efficiency
DeepSeek-V3.5 launches with 671B sparse MoE and Multi-Head Latent Attention.
- What happened
- DeepSeek AI released DeepSeek-V3.5, featuring 671B total parameters with 37B active parameters per token across 256 fine-grained routed experts and Multi-Head Latent Attention v2.
- Why it matters
- Compresses per-token KV cache memory by 93.3% relative to Multi-Head Attention, allowing 10x larger serving concurrency on identical GPU nodes and driving first-party API pricing down to $0.14 per 1M input tokens.
- Who is affected
- AI engineering leads, enterprise software architects, cloud infrastructure planners.
- Recommended action
- Benchmark internal reasoning and agent tasks against DeepSeek-V3.5 and evaluate FP8 self-hosted inference for high-volume pipelines.
Federal Frontier AI Regulation & Compliance
US Department of Commerce and NIST establish binding 10^26 FLOPs reporting rules.
- What happened
- The Bureau of Industry and Security (BIS) codified 15 CFR Part 702 under the Defense Production Act, mandating 15-day notification and red-team vulnerability filings for training runs exceeding 10^26 FLOPs or clusters of >=50,000 GPUs.
- Why it matters
- Transitions frontier AI oversight from voluntary commitments to statutory compliance backed by federal civil penalties of up to $250,000 per violation day.
- Who is affected
- Frontier AI labs, cloud hyperscalers, compliance officers, enterprise risk directors.
- Recommended action
- Implement computational run tracking, audit multi-tenant foreign compute leases, and prepare standardized red-teaming reporting protocols.
Semiconductor Manufacturing & Hardware Packaging
TSMC 2nm GAAFET yields reach 68.2% while CoWoS-L packaging books out through 2027.
- What happened
- TSMC Fab 20 disclosures confirmed 2nm GAAFET trial yields reached 68.2%, but advanced packaging lines (CoWoS-L and SoIC) are 100% committed through Q4 2027, with Nvidia taking 52% and Apple 26%.
- Why it matters
- Solidifies advanced packaging substrates and interposers as the primary physical choke point governing AI accelerator deliveries through 2027.
- Who is affected
- Semiconductor procurement leads, datacenter strategists, cloud infrastructure teams.
- Recommended action
- Plan 2027 infrastructure capacity based on packaging allocations rather than raw wafer supply and prioritize memory-efficient architectures like MLA.
Inference Infrastructure Security
Speculative Leak exploit exposes KV-cache side channels in multi-tenant inference.
- What happened
- Researchers from ETH Zurich and UC Berkeley demonstrated CVE-2026-38192, a microarchitectural cache-timing attack reconstructing prompt text with 91.4% accuracy across shared GPUs running speculative decoding.
- Why it matters
- Demonstrates that speculative decoding optimizations leak memory bus access patterns, requiring constant-time verification that reduces inference throughput by 18% to 24% unless physical MIG partitioning is enforced.
- Who is affected
- CISOs, cloud ML platform engineers, multi-tenant hosting providers.
- Recommended action
- Enforce hardware-level GPU isolation (MIG) for sensitive enterprise workloads or apply constant-time verification patches in speculative inference runtimes.
Since 2026-09-14
What changed
- DeepSeek released DeepSeek-V3.5, achieving 82.4% on AIME 2026 and 49.6% on SWE-bench Verified at $0.14 per 1M input tokens. Source (opens in a new tab)
- The US Department of Commerce finalized binding reporting rules under 15 CFR Part 702 for models trained with >10^26 FLOPs. Source (opens in a new tab)
- TSMC Fab 20 confirmed 68.2% yields on 2nm GAAFET test chips, but reported that 2027 CoWoS-L packaging is 100% committed. Source (opens in a new tab)
- Security researchers disclosed Speculative Leak (CVE-2026-38192), extracting prompt text via cache side channels on shared GPUs. Source (opens in a new tab)
- Enterprise infrastructure audits established the on-premises inference crossover point at 40 million tokens per day. Source (opens in a new tab)
Decision context
Why it matters
- Multi-Head Latent Attention (MLA) breaks the inference memory wall by compressing KV-cache footprints by 93.3%, unlocking dramatic serving cost reductions. Source (opens in a new tab)
- Frontier AI oversight has shifted from voluntary White House pledges to legally binding statutory requirements with $250,000/day penalties. Source (opens in a new tab)
- Advanced packaging substrates and thermal interposers—not raw wafer lithography—dictate global frontier compute availability through 2027. Source (opens in a new tab)
- Speculative decoding speedups in shared multi-tenant clouds come with microarchitectural side-channel risks that require dedicated physical isolation. Source (opens in a new tab)
- At sustained volumes above 40M tokens/day, private on-premise clusters running open-weight FP8 models reduce enterprise inference spend by 65%. Source (opens in a new tab)
Action and watchlist
What to do or monitor next
- Adoption of Multi-Head Latent Attention (MLA) across other open and commercial model architectures. Source (opens in a new tab)
- First formal filing submissions and disclosure audits under the BIS 10^26 FLOPs reporting mandate. Source (opens in a new tab)
- TSMC packaging capacity expansion timelines at Chiayi Advanced Packaging Fab 1. Source (opens in a new tab)
- Runtime security patches and constant-time speculative decoding updates across vLLM and TensorRT-LLM. Source (opens in a new tab)
No material change in other tracked categories
- Commercial US frontier proprietary rates remained unchanged, widening the pricing spread against open-weight sparse architectures.
Technical change log
Model, price, hardware and open-model movement
| Provider | Model | Availability | Modality | Best fit | Source |
|---|---|---|---|---|---|
- First-party frontier inference price floor dropped to $0.14 per 1M input tokens and $0.028 per 1M cached tokens for DeepSeek-V3.5.
- Enterprise infrastructure models identified an on-premises cost crossover at 40 million tokens/day, yielding a 65% cost reduction.
- TSMC confirmed 68.2% 2nm GAAFET trial yields at Fab 20 while reporting that 2027 CoWoS-L and SoIC packaging capacity is 100% booked.
- DeepSeek-V3.5 was released with open weights and full permissive inference support.
Benchmarks
Verified benchmark changes
- DeepSeek-V3.5 established open-weight frontier scores on AIME 2026 (82.4%), MATH-500 (96.8%), and SWE-bench Verified (49.6%).
Markets
August 13, 2026 United States market close
Tracked daily movement
Quote timestamp: 2026-08-13T16:00:00-04:00.
| Item | Value |
|---|---|
| SPX | +0.65% |
| DJI | +0.13% |
| IXIC | +0.81% |
| Ticker | Company | Close | Change | Source |
|---|---|---|---|---|
| SPX | S&P 500 | $7798.99 | +0.65% | Historical quote (opens in a new tab) |
| DJI | Dow Jones Industrial Average | $53839.99 | +0.13% | Historical quote (opens in a new tab) |
| IXIC | Nasdaq Composite | $26803.03 | +0.81% | Historical quote (opens in a new tab) |
Regular-session snapshot. Informational only; not investment advice.
Industry and policy
Professional context
DeepSeek Deploys DeepSeek-V3.5 with 671B MoE and MLA v2
DeepSeek-V3.5 achieves top-tier mathematical and coding performance while dropping token costs to $0.14/M input and $0.28/M output.
High impactOriginal source (opens in a new tab)Reviewed, corrected and approved by H. Omer Aktas.US Commerce Department Finalizes Dual-Use 10^26 FLOPs Rule
Codifies mandatory 15-day reporting and cybersecurity disclosure under the Defense Production Act with $250,000/day civil penalties.
High impactOriginal source (opens in a new tab)Reviewed, corrected and approved by H. Omer Aktas.TSMC Advanced Packaging 100% Committed Through 2027
Despite 68.2% yields on 2nm GAAFET silicon, CoWoS-L and SoIC capacity limits cap global AI accelerator expansion.
High impactOriginal source (opens in a new tab)Reviewed, corrected and approved by H. Omer Aktas.Inference Workloads Hit 40M Token On-Premises Crossover
Continuous enterprise inference at >40M tokens/day yields 65% cost savings on private 8-way H200/MI300X clusters vs public APIs.
Medium impactOriginal source (opens in a new tab)Reviewed, corrected and approved by H. Omer Aktas.Speculative Leak Microarchitectural Vulnerability Disclosed
CVE-2026-38192 reveals memory cache side-channel risks in speculative decoding engines on shared multi-tenant GPUs.
High impactOriginal source (opens in a new tab)Reviewed, corrected and approved by H. Omer Aktas.Speculative Leak: Microarchitectural Side-Channels in Speculative Decoding (CVE-2026-38192)
ETH Zurich and UC Berkeley document 91.4% prompt reconstruction accuracy via memory cache timing in shared GPU multi-tenancy.
Original source (opens in a new tab)Reviewed, corrected and approved by H. Omer Aktas.US BIS Enforces Mandatory Red-Teaming Reports for 10^26 FLOPs Models
Mandates quarterly technical filings on cyber-offense capabilities and weight protection safeguards under 15 CFR Part 702.
Original source (opens in a new tab)Reviewed, corrected and approved by H. Omer Aktas.Multi-Token Latent Attention Verification and Cache Memory Isolation
Analysis of low-rank KV-cache compression demonstrates resilience against uncompressed memory injection attacks.
Original source (opens in a new tab)Reviewed, corrected and approved by H. Omer Aktas.Limitations and unavailable information
- DeepSeek-V3.5 API pricing reflects first-party foundry availability; third-party cloud hosting rates range between $0.22/M and $0.38/M.
- BIS 15 CFR Part 702 reporting compliance begins 15 days post-publication; initial enforcement filings will be audited in subsequent quarters.
- Speculative Leak side-channel exploit requires non-MIG shared GPU multi-tenancy; physical hardware slicing (MIG) mitigates the vulnerability.
Audit appendix
How this edition was verified
The sections below are intended for readers who need publication controls, field-level history and traceability. They are separated from the default morning briefing.
Verified day-over-day comparison
What changed since 2026-09-14
2 material changes detected across five tracked lanes.
New, removed or materially revised model records.
20 current records trackedEndpoint, region, alias, access and lifecycle changes.
20 current records trackedAPI token prices, paid-plan terms and published promotions.
30 current records trackedComparable score, rank, coverage or methodology-status changes.
57 current records trackedPublished free-plan availability, limits and eligibility terms.
2 current records trackedMaterial movement only
Field-level changes
DeepSeek DeepSeek-V3.5
DeepSeek DeepSeek-V3.5 added to tracking
DeepSeek DeepSeek-V3.5
DeepSeek DeepSeek-V3.5 price or plan terms added to tracking
Unchanged lanes
- Availability: no material field change detected.
- Benchmarks: no material field change detected.
- Free tiers: no material field change detected.
Comparison method: Field-level day-over-day comparison. Source-link maintenance by itself is ignored, so a citation refresh cannot create a false product change.
Historical intelligence
Verified trend windows
Only preserved field-level changes are counted. Missing dates are never invented.
2 of 7 calendar days represented by 2 preserved editions
- Models
- 1
- Prices
- 1
- Benchmarks
- 0
2 of 30 calendar days represented by 2 preserved editions
- Models
- 1
- Prices
- 1
- Benchmarks
- 0
2 of 90 calendar days represented by 2 preserved editions
- Models
- 1
- Prices
- 1
- Benchmarks
- 0
Governed pricing intelligence
Pricing changes and source health
2 preserved editions from 2026-09-14 through 2026-09-15. Currencies and regions are never silently merged.
- API rates · introducedDeepSeek DeepSeek-V3.5
DeepSeek DeepSeek-V3.5 entered the tracked pricing database.
Source (opens in a new tab)
Source reliability and publication governance
Publication blocked
342 sources assessed · 34 used for critical claims · overall grade B (88/100).
- Expired for this evidence category
- Critical evidence grade D is below the publication threshold.
- Expired for this evidence category
Claim-level traceability
Citation coverage
Consequential statements and numerical values are mapped to explicit evidence instead of relying on page-level source lists.
4 claims require attention. Open the register to review weak, unsupported or invalid evidence.
Open the claim register →Correction integrity
Correction and revision ledger
No corrections or retractions are recorded for this edition. Future revisions must preserve the original value, replacement value, reason, affected pages, evidence and approval.
Open the complete correction ledger →Traceability
Sources used in this edition
- DeepSeek-V3.5 Technical Report: Multi-Head Latent Attention v2, 671B Sparse MoE Architecture and Inference Economics (opens in a new tab)DeepSeek AI · Primary technical release report and model card · Published 2026-09-15 · Retrieved 2026-09-15T09:00:00-04:00
- Bureau of Industry and Security Rule: Mandatory Reporting Requirements for Frontier Dual-Use Foundation Models Exceeding 10^26 FLOPs (15 CFR Part 702) (opens in a new tab)US Department of Commerce & NIST · Official federal regulatory notice and statutory compliance filing · Published 2026-09-15 · Retrieved 2026-09-15T09:00:00-04:00
- TSMC Hsinchu Fab 20 2nm GAAFET Production Update and 2027 Advanced Packaging CoWoS-L Allocation Audit (opens in a new tab)TSMC / Taiwan Ministry of Economic Affairs · Foundry technical disclosures and regulatory filings · Published 2026-09-15 · Retrieved 2026-09-15T09:00:00-04:00
- Speculative Leak: Microarchitectural Side-Channels and Cache Contention in Speculative Decoding Inference Engines (CVE-2026-38192) (opens in a new tab)ETH Zurich & UC Berkeley Systems and Security Lab · Academic cybersecurity research paper and coordinated vulnerability advisory · Published 2026-09-15 · Retrieved 2026-09-15T09:00:00-04:00
- Quarterly Enterprise AI Infrastructure Audit: The 40 Million Token Crossover from Public Cloud APIs to Dedicated On-Premises Clusters (opens in a new tab)Enterprise Cloud Infrastructure Research Group · Primary financial and cloud infrastructure telemetry report · Published 2026-09-15 · Retrieved 2026-09-15T09:00:00-04:00
Verification
Publication controls require attention
- Sources
- 342
- Evidence grade
- B
- Critical citations
- 97%
- Numerical citations
- 96%
- Corrections
- 0
- Blockers
- 10