2026.04.29 AI News Briefing — OpenAI joins AWS within 24 hours; DeepSeek's 90% cache price-cut shock

The AI cloud and pricing map is being redrawn within a single day. The day after announcing the end of its Microsoft exclusivity, OpenAI immediately launched its models, Codex, and a dedicated agent service on AWS Bedrock. At the same time, China's DeepSeek cut its 'cache hit' input price to one-tenth, and Musk took direct aim at OpenAI in a California courtroom, warning that 'AI could kill us all'.

---

1. OpenAI officially lands on AWS Bedrock — a multicloud move within 24 hours

• Source: OpenAI / AWS / Reuters / TechCrunch | 2026.04.28
• What happened: Just one day after announcing the end of its Microsoft exclusivity, OpenAI officially announced an expanded partnership with Amazon's AWS. GPT-5.4 is available immediately on AWS Bedrock, with GPT-5.5 to follow within weeks. The coding agent 'Codex' and a new service, 'Bedrock Managed Agents powered by OpenAI', launched alongside it. AWS CEO Matt Garman called it 'an announcement customers have waited a long time for', and Sam Altman, tied up in the courtroom battle with Musk, joined by video message.
• Why it matters: It is the decisive scene of OpenAI's transformation from 'model vendor' to 'infrastructure-neutral platform'. For Korean companies, the same OpenAI models can now be used on Azure, AWS, or Oracle, greatly reducing the burden of cloud lock-in.

2. DeepSeek cuts 'cache hit' input price by 90% — V4-Pro gets a 75% discount promotion

• Source: DeepSeek API Docs / Dataconomy / Investing.com Korea | 2026.04.26–27
• What happened: From April 26 (UTC), China's DeepSeek cut the 'input cache hit' price for all its models to one-tenth of the previous level. Note that the cut applies not to overall input and output prices but only to the 'cache hit rate charged when reusing a prompt prefix that has already been processed'. For V4-Flash, the cache hit rate is about $0.028 per million tokens, while cache-miss input is priced separately at $0.14 and output at $0.28. The new V4-Pro model has a separate 75% discount promotion running until May 5. DeepSeek says it runs V4 on Huawei Ascend chips and is passing the compute savings through to prices.
• Why it matters: For RAG and agent workloads with large system prompts and shared context, input cost effectively drops to 'rounding error' levels. Korean companies' bargaining power on LLM API pricing has gone up a notch, and US Big Tech now faces pressure to readjust its pricing.

3. Musk vs. Altman, day one in Oakland federal court — the frontal attack: "they stole a charity"

• Source: NBC News / CNN / NPR / Fortune / CNBC | 2026.04.28
• What happened: Elon Musk took the stand for the first time in federal court in Oakland, California. In opening statements, Musk's attorney Steven Molo took direct aim, saying the defendants 'stole a charity', and Musk testified that 'the reason OpenAI exists is that Larry Page called me a specieist' and that 'we want the Star Trek ending, not the Terminator ending'. Musk is demanding that OpenAI revert to nonprofit status, that Altman and Brockman be removed as directors and officers, and disgorgement of roughly $130 billion. OpenAI's attorney William Savitt countered that 'Musk simply left because he couldn't get his way at OpenAI'. Microsoft is also named as a co-defendant, and CEO Satya Nadella and others are expected to testify later. The trial is expected to run about four weeks.
• Why it matters: It is a precedent-setting case that will decide the legality of the 'nonprofit to for-profit' conversion. Depending on the verdict, OpenAI's IPO, governance, and IP licensing structure could be shaken, with direct implications for how later AI labs design their governance.

4. Google signs a new AI contract with the US Department of Defense — filling the seat Anthropic declined

• Source: LLM Stats | 2026.04
• What happened: Right after Anthropic refused to let the US Department of Defense (DoD) use its AI for 'domestic mass surveillance' and 'autonomous weapons', Google is reported to have signed a new contract with the DoD. The policies of individual AI labs on access to military and government markets are diverging sharply.
• Why it matters: AI safety and ethics policy is becoming a criterion that segments the market. When Korea's defense and public sectors adopt AI, reviewing each model's 'use policy' looks set to become a de facto required step.

5. AWS unveils a full agent SaaS lineup — expanding the front with four Quick and Connect products

• Source: SiliconANGLE / The Register / The NAS Guy | 2026.04.28
• What happened: Alongside the OpenAI announcement, AWS also unveiled a full lineup of its own agent SaaS products. The desktop AI assistant 'Amazon Quick' launched with free and Plus plans, and 'Amazon Connect' expanded into four solutions: Decisions (supply chain), Talent (recruiting), Customer (CX), and Health (healthcare). CEO Garman said that 'Prime Video used the new agent solution to finish a code rewrite that would have taken two years in two quarters'.
• Why it matters: In the middle of the 'SaaSpocalypse' debate, AWS has offered its answer: 'we are not SaaS, we are agent infrastructure'. In martech, DAM, and CX, tool selection criteria are shifting from 'features' to 'agent integration'.

6. Korea launches the 'AI Privacy Public-Private Policy Council' — targeting agentic and physical AI

• Source: Korea Policy Briefing / Personal Information Protection Commission | launched 2026.02, activity ramping up in April
• What happened: The government launched the '2026 AI Privacy Public-Private Policy Council'. To respond to the spread of agentic AI and physical AI, it operates three subcommittees: data processing, risk management, and data subject rights. It is intended to keep pace with global regulatory trends such as the EU AI Act, which takes full effect in August, and the Colorado law. In Korea, documenting 'AI model cards' and 'data lineage' is expected to become the audit standard soon.
• Why it matters: It is the stage where the compliance burden for Korean companies adopting AI becomes concretely visible. Content governance systems such as AEM and DAM are likely to gain 'AI usage tracking' as a required feature.

---

Today's insight

• The era of 'single cloud' AI infrastructure is over: As the 24-hour sequence of Microsoft–OpenAI exclusivity ending (yesterday) followed by immediate AWS–OpenAI integration (today) shows, models now move freely across clouds. Korean companies should define 'which governance layer' before 'which cloud'.
• 'Cache hit' pricing is the new front: DeepSeek's cut is not a simple list-price reduction but a structural change that takes 'repeated prompts to nearly free'. The larger the system prompt and context, as in RAG and agent workloads, the more extreme the price gap becomes. Price readjustments from OpenAI, Anthropic, and Google are likely within the next six months.
• AI ethics and policy cannot be separated from commercial strategy: Anthropic's DoD refusal and Google's contract, the Musk–Altman courtroom battle, and the launch of Korea's AI privacy council are all different facets of the same trend. Per-model use policies and governance design now decide market access.